Top 10 Open Source LLMs in 2026: DeepSeek V4, Kimi & More

The definitive July 2026 ranking of open-weight AI models: verified benchmarks, real API pricing, license fine print, and what actually changed since DeepSeek started this revolution.

GLM-5.2 scores 51 on the Artificial Analysis Intelligence Index. Claude Opus 4.8, the strongest closed model in the world, scores roughly 55.7. That gap, about four and a half points, is the entire distance between the best AI money can buy and the best AI you can download for free - Artificial Analysis. Eighteen months ago that gap was a chasm. Today it is a rounding error for most workloads, and the models on the open side cost 10x to 150x less to run.

But here is the problem: almost every "top open source LLMs" list you will find, including the previous version of this very guide, is describing a world that no longer exists. DeepSeek V3.2 has been superseded by DeepSeek V4. Meta killed the Llama line and went proprietary. Llama 3, Phi-3, Falcon, and StarCoder no longer appear in any serious 2026 ranking. The legacy deepseek-chat and deepseek-reasoner model names retire on July 24, 2026, sixteen days from this update - DeepSeek API Docs. If your model list is more than a quarter old, it is not a list, it is an archaeology exhibit.

This guide is the fully rebuilt July 2026 edition. It covers the top 10 open-weight models ranked with a transparent weighted methodology, the verified benchmark numbers that matter now (SWE-bench Verified, Terminal-Bench 2.1, the agentic-weighted AA Index v4.1), real API pricing down to cache-hit economics, the license fine print that determines whether you can actually ship, the geopolitical inversion that put Chinese labs on top of open AI, and a practical framework for choosing a model without getting whiplash from monthly deprecations. It is written for people who make decisions about AI, not just people who train models.

Contents

1. What Changed Since Late 2025
2. Benchmarks That Matter in 2026
3. DeepSeek V4: Pro and Flash
4. GLM-5.2: The Open-Weights Intelligence Leader
5. Kimi K2.6: The Agentic Trillion-Parameter MoE
6. MiniMax M3: Long-Context Economics
7. Qwen3.5: The Full-Stack Model Family
8. gpt-oss-120b: The Most-Downloaded Big Open Model
9. Gemma 4: Small Models, Serious Multimodality
10. Xiaomi MiMo-V2.5-Pro: The Trillion-Scale Dark Horse
11. Mistral Large 3: Europe's Apache 2.0 Flagship
12. Nemotron 3 Ultra and OLMo 3: The US Rearguard
13. Meta's Exit: How the Llama Era Ended
14. China Now IS Open-Source AI
15. The Frontier Gap, Quantified
16. Licenses Decoded: MIT, Apache 2.0, and Display Clauses
17. Pricing Economics: Cache Hits and Cost Per Task
18. Context Windows: 1M Tokens Is the New Baseline
19. Self-Hosting Hardware Tiers in July 2026
20. Deprecation Watch and Migration Guidance
21. From Models to Agents: Putting Open Weights to Work
22. Conclusion: Choosing Your Model in July 2026

The July 2026 Scorecard

Before the deep dives, here is the master ranking. Every model below is scored on four criteria weighted by what actually determines success when you deploy an open model: raw intelligence and agentic capability (35%), cost efficiency (25%), deployability covering license, context window, and hardware reality (20%), and ecosystem adoption measured by downloads, provider availability, and tooling (20%). Scores run 0-10, and every cell shows the evidence behind the number, not just the number.

The ranking rewards the model you would actually build on, which is why the AA Intelligence Index leader does not automatically take the top slot. Intelligence is one input among several. The winner is the model where intelligence, price, license, and ecosystem all compound rather than trade off against each other.

Sort check: 9.3, 8.7, 8.4, 8.4, 8.2, 8.0, 7.8, 7.6, 7.5, 7.3, 6.8. Descending order confirmed. Kimi K2.6 and MiniMax M3 tie at 8.4 and are ordered alphabetically. The scores and evidence come from the Artificial Analysis open-weights leaderboard, official model cards, and each lab's published pricing, all as of July 2026. Now let us get into why this table looks nothing like it did seven months ago.

1. What Changed Since Late 2025

The previous edition of this guide, published in December 2025, ranked DeepSeek V3.2 first, Llama 3 70B second, and included Phi-3, Baichuan 2, H2O GPT, Falcon, and StarCoder in the top ten. Every single one of those picks is now obsolete, and understanding why is more useful than any individual ranking, because the same forces will eventually obsolete this list too. The structural fact underneath everything: open-weight AI now ships on a three-to-four-month cadence, and each generation does not just improve on the last, it replaces it commercially, with old API names deprecated within a quarter.

Three events defined the last seven months. First, DeepSeek V4 arrived on April 24, 2026 as a two-model family, with 1M-token context as the default across all official DeepSeek services and a hard retirement date for the old model names - DeepSeek API Docs. Second, Meta exited open weights entirely: instead of a Llama 5, Meta Superintelligence Labs shipped the proprietary Muse Spark in April 2026, inverting the article-of-faith that Meta was open-source AI's patron - Meta AI. Third, the leaderboard crown changed hands twice, with Z.ai's GLM-5 (February) and then GLM-5.2 (June 13, 2026) taking the top open-weights score on the Artificial Analysis Intelligence Index - GLM-5 GitHub.

The velocity is the story. In one seven-month window, five different labs shipped models that would each have been the undisputed best open model on Earth a year earlier. The December 2025 assumption that a 128K context window was generous now looks quaint: 1M tokens is the default across DeepSeek V4, GLM-5.2, MiniMax M3, and MiMo-V2.5, and Llama 4 Scout still holds the outlier record at 10M - Meta AI. The old "240% enterprise adoption growth" statistics of 2023-2025 vintage have been replaced by a harder number: open-weight models now carry roughly one-third of all token volume on OpenRouter, the largest neutral model marketplace - OpenRouter.

Notice also which names disappeared without a formal announcement. Falcon and StarCoder, pioneers of 2023-2024 open AI, simply stopped being competitive: StarCoder's 15B parameters and 8K context read like specifications from a different technological era next to a 1M-context trillion-parameter MoE. Phi-3, once celebrated for matching early GPT-3.5, was superseded inside Microsoft's own lineup by Phi-4, and the entire "small model that punches up" category it defined now belongs to Gemma 4 and the Qwen small tier. Obsolescence in this market rarely arrives as a press release; it arrives as silence while the leaderboards move on.

For anyone who built on the previous generation, the practical takeaway is blunt. Models are now infrastructure with deprecation schedules measured in months: Xiaomi discontinued its commercial MiMo-V2 series on June 30, 2026, and DeepSeek's legacy names go dark July 24. We cover the migration playbook in section 20, but the mindset shift comes first: you are not choosing a model, you are choosing a lab's release train, and you should evaluate the train's schedule as carefully as the current car.

2. Benchmarks That Matter in 2026

The previous version of this guide benchmarked models against MMLU scores and made comparisons to "OpenAI's early GPT-3.5." That framing is dead, and it is worth spending a section on why, because reading 2026 model cards with 2024 benchmark literacy leads to genuinely bad purchasing decisions. The structural change is that models stopped being chat engines and became agents: systems that operate terminals, browse the web, edit real codebases, and run for hours. Static multiple-choice knowledge tests saturate near 90% for every serious model, so they no longer discriminate between candidates.

The industry's composite reference has followed suit. The Artificial Analysis Intelligence Index v4.1 now combines nine evaluations explicitly weighted toward agentic work: GDPval-AA v2, tau3-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR - Artificial Analysis. When this guide cites an "AA score," that is the number in question: a weighted composite where operating a bank's back office and surviving a real terminal session count for more than trivia recall. We maintain a broader reference on evaluation methodology in our AI model benchmarks and pricing guide if you want the full landscape.

Beyond the composite, four individual benchmarks do most of the work in 2026 model selection. Each measures a distinct failure mode, which is why serious model cards now report all of them rather than cherry-picking one.

• SWE-bench Verified: resolves real GitHub issues in real repositories; the standard for coding agents
• SWE-bench Pro: the harder, contamination-resistant successor; scores drop 15-25 points versus Verified
• Terminal-Bench 2.1: multi-step tasks in a live shell; the best proxy for autonomous operations work
• BrowseComp: agentic web research requiring long multi-site browsing chains

The reason these four matter is economic, not academic. A model that scores 80% on SWE-bench Verified can close four of every five real bug tickets you hand it, which converts directly into engineering payroll math. A model that scores well on MMLU but poorly on Terminal-Bench will happily explain what a Kubernetes pod is and then destroy one when given actual access. The gap between Verified and Pro scores also functions as a contamination detector: a model whose Verified score towers over its Pro score may have seen the easier benchmark's problems during training, which is why this guide reports both for the models that publish them.

One more calibration note before the rankings: benchmark snapshots from different sources can disagree, and honest reporting means saying so. OpenRouter's June 2026 analysis credits NVIDIA's Nemotron 3 Ultra with an AA-style score of 48, while the July Artificial Analysis leaderboard shows a Chinese sweep of the top six - OpenRouter. Index versions, evaluation dates, and reasoning-mode settings ("Think Max" versus standard) all move scores by points. Treat every number in this guide as a verified snapshot with a source, not a physical constant. With the measuring instruments calibrated, we can now look at the models themselves, starting with the lab that gave this revolution its name.

3. DeepSeek V4: Pro and Flash

DeepSeek remains the center of gravity in open-weight AI, and the V4 generation, released April 24, 2026, is the most complete argument for that position yet. The family comes in two Mixture-of-Experts variants: DeepSeek-V4-Pro at 1.6 trillion total parameters with 49B active per token, and DeepSeek-V4-Flash at 284B total / 13B active - DeepSeek API Docs. Both ship with 1M-token context as the default, both are open-weight, and the Pro model is MIT licensed, the least restrictive mainstream license in software. For the full origin story of this release, see our DeepSeek V4 launch guide published the day it dropped.

The engineering under the hood explains the economics. V4-Pro uses FP4+FP8 mixed precision and a hybrid Compressed Sparse Attention scheme, which is how a 1.6T-parameter model serves a million tokens of context without a million-dollar serving bill - Hugging Face. The benchmark results land at the top of the open cohort: 80.6% SWE-bench Verified, 93.5 LiveCodeBench, a Codeforces rating of 3206 (grandmaster territory), and 87.5 MMLU-Pro in Think Max mode. On the AA Intelligence Index, V4 Pro scores 44 and Flash scores 40, placing both variants inside the open top six. Adoption is commensurate: roughly 1.2M Hugging Face downloads in the past month for Pro alone.

Pricing is where DeepSeek breaks the industry's brain, and the official numbers deserve a table - DeepSeek pricing:

Read that cache-hit number again: $0.0028 per million input tokens. An agent that repeatedly consults the same million-token codebase pays about a quarter of a cent per full-context read after the first pass. OpenRouter's June 2026 analysis found discount providers serving Flash at $0.054/$0.242 per million, roughly 150x cheaper than GPT-5.5 output pricing, while Flash scores 79% on SWE-bench Verified - OpenRouter. Section 17 works through what this does to cost-per-task math, but the one-line version: for continuous agentic workloads, Flash is not merely cheaper than the closed frontier, it is cheaper than most companies' logging bills.

The two-variant strategy itself is worth understanding because it will be copied. Pro exists to contest the intelligence frontier; Flash exists to win the volume war, and each makes the other more credible. A team can prototype against Pro, confirm the workload works, then route the 90% of traffic that does not need peak intelligence to Flash at one-third the price, all inside one lab's ecosystem, one API shape, one set of quirks. This frontier-plus-workhorse pairing is the same product logic closed labs use across their model tiers, executed with open weights and MIT terms.

The caveats are real but narrow. The V4 release came with a hard deprecation: deepseek-chat and deepseek-reasoner, the API names half the industry hardcoded in 2025, retire on July 24, 2026. And V4-Pro's 1.6T total parameters put true self-hosting out of reach for all but multi-node clusters, which is why most teams consume it through providers. Neither caveat dents the conclusion: on the combined weight of intelligence, price, license, and ecosystem, DeepSeek V4 is the #1 open model family of July 2026. What it is not, any longer, is the single smartest open model. That title moved to Beijing's other lab.

4. GLM-5.2: The Open-Weights Intelligence Leader

If the question is purely "what is the smartest model I can download," the July 2026 answer is GLM-5.2 from Z.ai (formerly Zhipu AI). Released June 13, 2026, it scores 51 on the AA Intelligence Index, the highest mark ever recorded by open weights and within about 4.7 points of Claude Opus 4.8's ~55.7 - Artificial Analysis. The previous edition of this guide covered GLM-4.6 at 355B parameters and a 200K context window; three versions and seven months later, GLM-5.2 is a 744B total / 40B active MoE with a 1M-token context, released under Apache 2.0.

The generational leap was deliberate and heavily resourced. GLM-5, the February 11, 2026 base release, was pretrained on 28.5 trillion tokens, more than doubling the parameter count of its GLM-4.5 ancestor - GLM-5 GitHub. GLM-5.1 followed on April 8 and GLM-5.2 on June 13, a three-release cadence in four months that illustrates how compressed open-model iteration has become. The agentic numbers are the headline: 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro, the contamination-resistant benchmark where every model's score deflates. That Terminal-Bench figure means GLM-5.2 completes four of five realistic multi-step shell tasks, which is the single best predictor of usefulness for autonomous operations.

Speed and economics keep pace with intelligence. Artificial Analysis measures GLM-5.2 serving at 215 tokens per second, unusually fast for a model this large, and Z.ai markets coding-agent sessions at roughly $0.26 per task, about one-sixth the cost of running the same workload through GPT-5.5. We published a dedicated GLM-5.2 practical guide and a separate benchmarks and cost breakdown if you are evaluating it seriously.

Z.ai's trajectory also illustrates the new capital structure behind Chinese open labs. The company completed a Hong Kong IPO on January 8, 2026, reportedly raising around US$558M, a figure worth treating as secondary-source until confirmed against filings, and the release cadence since suggests the proceeds went straight into training runs. A lab that answers to public markets has both more fuel and more pressure to hold the leaderboard crown, which for buyers cuts both ways: faster improvements, but also faster deprecation of anything that stops earning its serving costs.

Why does GLM-5.2 rank second here despite the top intelligence score? Weighting. Its 744B total parameters make self-hosting heavier than DeepSeek's Flash tier, its ecosystem of providers, fine-tunes, and tooling is younger than DeepSeek's, and its pricing, while excellent, does not reach Flash's cache-hit absurdity. None of that diminishes the achievement. For maximum downloadable intelligence per query, GLM-5.2 is the pick, full stop. For maximum intelligence per dollar across a whole agent fleet, the table's #1 keeps its crown, and the next two entrants tie each other trying to take it.

5. Kimi K2.6: The Agentic Trillion-Parameter MoE

Moonshot AI was entirely absent from the previous edition of this guide, which now reads as its biggest omission. The Beijing lab's Kimi K2.6, released April 2026, is a 1 trillion total / 32B active MoE with 384 experts (8 selected per token), a 256K context window, and a native 400M-parameter MoonViT vision encoder, making it genuinely multimodal rather than text-first with vision bolted on - Hugging Face. Its benchmark card is elite: 80.2% SWE-bench Verified, 96.4% AIME 2026, 90.5% GPQA-Diamond, and 89.6 LiveCodeBench, with an AA Index score of 44 tying DeepSeek V4 Pro.

The adoption numbers are arguably more impressive than the benchmarks. Kimi K2.6 logged roughly 1.99M Hugging Face downloads in the past month, second only to gpt-oss-120b among large open models and well ahead of DeepSeek V4 Pro. That reflects something real about how the model behaves in practice: K2.6 and its January predecessor K2.5 were built agent-first. K2.5 was continually pretrained on roughly 15 trillion mixed visual and text tokens on top of Kimi-K2-Base, and its Agent Swarm architecture coordinates up to 100 parallel sub-agents on a single task - Kimi K2.5 GitHub. We dissected the swarm economics in our Kimi K2.6 agent swarm guide, including when parallel sub-agents actually save money versus burn it.

The vision-native design deserves a concrete illustration, because "multimodal" has become an empty checkbox word. A MoonViT-equipped agent can watch a browser session, read the error state from a screenshot, cross-reference the failing element against the codebase in its context, and patch the front-end bug, all in one coherent loop with no separate OCR or captioning model gluing modalities together. Agents that operate real software spend most of their perception budget on screens, and a model that perceives screens natively wastes none of that budget on translation layers. This is the workload where K2.6's download numbers come from.

Licensing is the one nuance buyers must read rather than assume. Kimi ships under a Modified MIT license: standard MIT terms plus exactly one added condition, that products exceeding 100M monthly active users or $20M per month in revenue must display "Kimi K2" in their user interface - Kimi license. Below those thresholds it behaves as ordinary MIT. For the overwhelming majority of companies this clause will never trigger, but legal teams at consumer-scale platforms need to know it exists, and section 16 places it in the broader license taxonomy.

The practical positioning: choose Kimi K2.6 when your workload is agentic and multimodal at once, screenshots plus code plus long tool chains, and when 256K context suffices. Its context window is the shortest among the top four, native 256K versus the 1M standard, which is the main reason it does not rank higher. For pure long-document work, the next model on the list was built for exactly that.

6. MiniMax M3: Long-Context Economics

MiniMax is the second lab on this list that did not exist in the previous edition, and its June 1, 2026 release M3 is the purest expression of the 2026 thesis that context length and serving cost, not raw intelligence, are where open models compete hardest. M3 is a 428B total / 23B active MoE with a 1M-token context window and native image and video understanding, scoring 44 on the AA Intelligence Index, tied with DeepSeek V4 Pro and Kimi K2.6 - Artificial Analysis. Three models tied at 44 from three different labs is itself remarkable: the open frontier is now a pack, not a leader with stragglers.

M3's differentiation is MiniMax Sparse Attention, which delivers up to 15x faster decoding on long-context workloads. That number matters because 1M-token context windows are usually a marketing checkbox with a hidden tax: attention cost grows brutally with sequence length, so models technically support long contexts that nobody can afford to use at production volume. M3 attacks the tax directly, and pairs it with serving prices around $0.098 per million input tokens and $1.21 per million output, per OpenRouter's June 2026 provider data - OpenRouter.

The February predecessor M2.5 remains relevant and shows the lab's pricing philosophy. M2.5 scores 80.2% SWE-bench Verified and 76.3% BrowseComp, the latter being the best agentic web-research score in the open cohort, and its Lightning variant is priced at $0.3/M input, $2.4/M output at 100 tokens per second, which MiniMax markets with the memorable framing of "$1 to run the model continuously for an hour" - MiniMax. Standard M2.5 costs half that. Pricing an AI model like electricity, by the hour of continuous operation, tells you exactly which customer MiniMax is chasing: teams running always-on agents rather than occasional chat.

That BrowseComp score deserves a second look because web research is quietly one of the highest-value agent workloads in production. A 76.3% BrowseComp result means the model completes three of four long-horizon research tasks that require chaining searches, reading dozens of pages, reconciling contradictions, and synthesizing an answer. Combined with the 1M context that lets it hold every page it has read in working memory, M3 and its M2.5 sibling are arguably the strongest open stack for research automation, the category of work where token volume explodes and closed-frontier pricing hurts most.

Like Z.ai, MiniMax completed a Hong Kong IPO in January 2026, giving it public-market capital to sustain the release cadence (treat specific raise figures as secondary-source pending filings). The buying guidance is clean: M3 is the strongest choice when your workload is long-document plus multimodal, entire codebases, legal discovery sets, hours of video, and when you want frontier-adjacent intelligence at the lowest sustained serving cost in the 1M-context club. Where MiniMax offers one big model done well, the next entrant offers the opposite strategy: every size, one license.

7. Qwen3.5: The Full-Stack Model Family

Alibaba's Qwen was ranked third in the previous edition of this guide on the strength of Qwen 2.5 and early Qwen 3. The lab has since shipped two full generations. Qwen3.5-397B-A17B, released February 16, 2026, is the flagship: 397B total / 17B active parameters across 512 experts (11 active per token), Apache 2.0 licensed, with a vision encoder for image and video and coverage of 201 languages - Hugging Face. Its native context is 262,144 tokens, extensible to roughly 1,010,000 via YaRN scaling. Benchmarks put it solidly in the top tier: 76.4% SWE-bench Verified and 87.8% MMLU-Pro.

What makes Qwen unique is not the flagship but the family. The Qwen3.5 open-weight line spans 0.8B to 397B parameters across eight sizes, every one of them Apache 2.0, and the train did not stop: Qwen3.6-35B-A3B shipped April 16, 2026, followed by the dense Qwen3.6-27B on April 22 - Qwen3.6 GitHub. No other lab offers a coherent same-architecture ladder from a model that runs on a phone to a model that contests the frontier. For teams that prototype small and scale up, or distill large into small, that ladder is worth more than a few benchmark points.

The family structure has a compounding effect that single-model labs cannot match: everything transfers. A fine-tuning pipeline built for the 4B works on the 35B; a safety evaluation run against the 27B mostly holds for its siblings; a prompt library tuned on one size degrades gracefully rather than catastrophically when moved. Enterprises that standardize on Qwen are not standardizing on a model, they are standardizing on an architecture with eight price points, which is why Qwen remains the default fine-tuning substrate of the open ecosystem even when its flagship is not the leaderboard king.

The multilingual coverage deserves specific emphasis because it is a moat nobody else has matched. 201 languages is not a rounding-up of tokenizer coverage; Qwen has been the default open model for non-English deployment since 2024, and 3.5 extends the lead. If your users write Vietnamese, Swahili, or Kazakh, the practical model shortlist has one name on it. The vision stack similarly handles document parsing and video understanding natively, which matters for the enormous volume of enterprise work that is really "read this scanned PDF" in disguise.

Why fifth place, then? Two honest reasons. First, on the hardest agentic evals the flagship trails the 44-club: 76.4% SWE-bench Verified is excellent but measurably behind DeepSeek V4, Kimi, and GLM-5.2. Second, the native 262K context, while YaRN-extensible, is not the same as the trained-native 1M of DeepSeek V4 or MiniMax M3; long-context quality degrades faster on extended windows than native ones, a distinction section 18 unpacks. Neither weakness matters for the family's core role, and that role sets up an interesting contrast with the American model that everyone downloads but nobody's lab is iterating.

8. gpt-oss-120b: The Most-Downloaded Big Open Model

Here is the most counterintuitive data point in open AI: the most-downloaded large open model in July 2026 is made by OpenAI. gpt-oss-120b, released August 5, 2025 under Apache 2.0, logged roughly 4.3 million Hugging Face downloads in the past month, more than double Kimi K2.6 and triple DeepSeek V4 Pro - Hugging Face. An eleven-month-old model outdownloading the entire 2026 frontier tells you something important about the difference between what benchmarks measure and what deployment actually rewards.

The design explains the popularity. gpt-oss-120b is a 117B total / 5.1B active MoE that fits on a single 80GB GPU via MXFP4 quantization. That single-GPU fit is the magic property: one H100, A100, or workstation-class card runs the whole model with no cluster orchestration, no multi-node interconnect, no serving team. Its intelligence is solidly below the 2026 Chinese frontier, which is why it scores 6 on our intelligence criterion, but it clears the bar for the enormous middle of real workloads: summarization, extraction, internal copilots, moderate coding help.

The download gap visualized above is the ecosystem criterion made concrete: benchmark leadership and deployment leadership are different competitions with different winners. Three structural forces keep gpt-oss on top of the download charts. Hardware fit, as covered. Brand trust: compliance departments that hesitate over Beijing-lab weights approve an OpenAI artifact under Apache 2.0 without a meeting. And ecosystem inertia: eleven months of tutorials, fine-tunes, quantizations, and deployment recipes compound into the path of least resistance for every new team entering open AI.

The strategic question is whether gpt-oss gets a successor. OpenAI's commercial attention sits with its closed line, where GPT-5.5 anchored the frontier this spring and GPT-5.6 arrived in early July 2026. The open release looks increasingly like a strategic hedge rather than a product line, refreshed when regulatory or competitive pressure demands rather than on a roadmap. Until a successor lands, gpt-oss-120b remains the correct answer to a specific, common question: "what is the best model I can run on the one big GPU I already have, with a license nobody will question?" For hardware a notch below that, Google has the answer.

9. Gemma 4: Small Models, Serious Multimodality

Google occupies a strange position in open AI: its flagship intelligence ships closed in the Gemini 3.1 Pro line, while its open contribution, Gemma, targets the opposite end of the hardware spectrum and does it better than anyone. Gemma 4, released April 2, 2026, ships in five variants: E2B and E4B efficiency models for phones and edge devices, a 12B, a 26B-A4B MoE, and a dense 31B, all with audio and visual input, 140 languages, and native function calling - Google DeepMind.

The 31B thinking variant is the family's benchmark face, holding an Arena Elo of 1452 and 85.2% MMLU, remarkable output for 31B dense parameters. But the E-series is the strategically interesting part. A model that runs on a phone with audio and vision input is not competing with DeepSeek V4; it is competing with the cloud round-trip itself. On-device inference means zero marginal cost, zero latency floor, and data that never leaves the handset, three properties no API price can match. For consumer apps, wearables, field hardware, and privacy-mandated deployments, Gemma 4 E-series is effectively the only serious open option.

Native function calling in a small model is quietly the feature that unlocks agents at the edge. An E4B model on a phone that can reliably emit structured tool calls turns the device into an agent host: it can search local files, control apps, and hand off only the genuinely hard reasoning to a cloud model. That hybrid pattern, small open model as router and executor with a frontier model on call, is emerging as the dominant consumer architecture of 2026, and Gemma 4 was explicitly built for the local half of it.

Gemma's license is worth a note of caution relative to its peers: it ships under Google's custom Gemma terms rather than Apache 2.0 or MIT, with use-policy restrictions that most commercial teams find acceptable but that legal review should actually read. In exchange you get the deepest small-model tooling in the ecosystem: first-class support in every inference runtime, quantization-aware releases, and Google's distillation pipeline feeding capability down from Gemini into open weights. Our guide to cutting LLM costs covers when a 4B on-device model beats a 400B cloud model on total cost of ownership, and the answer is: more often than most teams assume.

The honest limitation is the ceiling. Nothing in the Gemma 4 family contests the agentic frontier; the 31B scores well for its size class and no further. Google's open strategy is to own the edge and the classroom, not the leaderboard. That leaves the top of the size spectrum to the labs that want it, and the next one on our list wants it badly enough to have trained a trillion-parameter model in FP8.

10. Xiaomi MiMo-V2.5-Pro: The Trillion-Scale Dark Horse

The most surprising name in the July 2026 top ten is a phone manufacturer. Xiaomi's MiMo-V2.5-Pro, released April 28, 2026, is a 1.02 trillion total / 42B active MoE under a clean MIT license, with context up to 1M tokens, trained on 27 trillion tokens in FP8 - Hugging Face. It scores 42 on the AA Intelligence Index, placing a company best known for smartphones and home appliances fifth among all open-weight labs on Earth, ahead of every American and European entrant except NVIDIA's best.

The architecture contributes a genuinely novel efficiency trick: a hybrid sliding-window plus global attention design that cuts KV-cache memory roughly 7x. KV-cache is the silent budget killer of long-context serving, the per-token memory that balloons as conversations grow, so a 7x reduction converts directly into either 7x more concurrent users per GPU or dramatically cheaper 1M-context sessions. Combined with FP8 training economics, MiMo-V2.5-Pro demonstrates that the trillion-parameter club is no longer gated on frontier-lab budgets, a first-principles shift: when consumer electronics margins can fund a frontier-adjacent model, the number of possible frontier labs stops being single-digit.

Xiaomi's motive is vertical integration, and buyers should read the roadmap accordingly. The company ships hundreds of millions of devices a year and wants its own intelligence stack from handset NPU to cloud, not a dependency on rivals. That explains both the aggressive open release, ecosystem building, and the ruthless deprecation: Xiaomi's commercial API discontinued the entire MiMo-V2 series on June 30, 2026, only months after launch, migrating everything to V2.5 - Xiaomi MiMo. Adoption is still early at roughly 101K Hugging Face downloads per month, two orders of magnitude behind gpt-oss, which is why its ecosystem score drags the final ranking despite top-five intelligence.

There is a broader lesson in Xiaomi's arrival that matters for anyone forecasting this market. The previous edition of this guide treated the open-model roster as a fixed cast: Meta, Mistral, Alibaba, a few research institutes. In twelve months, a phone maker, a video-AI startup (MiniMax), and a chatbot company (Moonshot) all shipped top-six models. The entry barrier is now capital plus efficiency engineering, not accumulated research pedigree, which means next year's list will contain names not on anyone's radar today. Plan model strategy for a market where the cast keeps changing.

The buying guidance: MiMo-V2.5-Pro is the value pick for teams that want 1M-context, MIT-licensed, top-five intelligence and are comfortable being early. The license is cleaner than Kimi's, the context is longer than Qwen's native window, and the KV-cache economics are the best in the trillion class. What it lacks is the surrounding ecosystem, which for many teams is precisely what they pay Europe's flagship lab to provide.

11. Mistral Large 3: Europe's Apache 2.0 Flagship

Mistral AI appeared in the previous edition of this guide via Mistral Small 3, a 24B dense model, with pricing quoted at $0.30 per million tokens. The company has since gone big. The Mistral 3 family, announced December 2, 2025, is headlined by Mistral Large 3: a 675B total / 41B active sparse MoE released under Apache 2.0 and trained on 3,000 NVIDIA H200 GPUs - Mistral AI. API pricing lands at $0.5 per million input tokens and $1.5 per million output, and the family extends downward through Ministral 3 dense models at 14B, 8B, and 3B, with the 14B reasoning variant hitting 85% on AIME '25. A Medium 3.5 refresh followed on April 29, 2026.

Judged purely on leaderboard position, Large 3 sits behind the Chinese frontier pack, which our intelligence score of 7 reflects. Judged on what European enterprises actually need to deploy AI, the calculus changes. Mistral is the only frontier-adjacent lab that is headquartered in the EU, contractable under EU law, and aligned with EU AI Act compliance timelines by default. For banks, insurers, healthcare groups, and public-sector buyers inside the regulatory perimeter, "almost as smart, fully Apache 2.0, and French" beats "smartest, MIT, and Chinese" in most procurement meetings, a dynamic section 14 examines from the other direction.

The Ministral line deserves independent attention rather than footnote status. An 8B or 14B dense model under Apache 2.0 with strong reasoning scores is the workhorse class for high-volume production: classification, routing, extraction, guardrails, first-draft generation. These tasks constitute most enterprise token volume, and paying frontier prices for them is the most common AI budgeting mistake of 2026. A Ministral 14B that hits 85% on AIME at a fraction of flagship cost is the right tool for that tier, and the same-family upgrade path to Large 3 removes the integration tax of mixing labs.

The 3,000-H200 training figure also tells a quieter story about efficiency convergence. Mistral built a 675B-parameter Apache 2.0 flagship on a cluster that would have been considered mid-sized for a frontier run in 2024, using the same sparse-MoE playbook as its Chinese competitors. The techniques that export controls forced Chinese labs to invent, aggressive sparsity, low-precision training, attention compression, have diffused globally within quarters of publication. Efficiency innovations do not stay proprietary in a field where weights, papers, and engineers all circulate, which is a structural reason to expect the open frontier to keep pace.

The strategic risk for Mistral is squeeze. Above it, Chinese labs iterate faster on raw capability; below it, Gemma 4 and Qwen's small models compete hard for the efficient tier. Its December 2025 flagship was state-of-the-art among open sparse models for weeks before the February-June wave rolled through. The company's bet is that sovereignty plus sufficiency is a durable moat even without the leaderboard crown, and given the direction of EU AI procurement policy, that bet looks sound. The final ranked entries test the mirror-image American version of the same thesis.

12. Nemotron 3 Ultra and OLMo 3: The US Rearguard

The strongest US-built open-weight model in July 2026 comes from NVIDIA. Nemotron 3 Ultra is a 550B total / 55B active hybrid combining Mamba-2 state-space layers with Transformer attention, and OpenRouter's June 2026 analysis credits it with a 48 on the AA Intelligence Index, the best score of any non-Chinese open model - OpenRouter. Serving prices run $0.423 per million input and $2.61 per million output with a free tier available. The output price is the catch: at nearly triple DeepSeek V4 Pro's $0.87, Nemotron is the rare open model that costs more than its Chinese benchmark peers, which caps its cost score in our table.

NVIDIA's motive differs from every other lab here, and understanding it predicts the roadmap. NVIDIA does not need model revenue; it needs reasons for the world to buy GPUs, and a first-rate open American model that showcases hybrid Mamba-Transformer serving efficiency on NVIDIA hardware is marketing, ecosystem lock-in, and a national-champion play at once. That makes Nemotron unusually likely to keep receiving investment regardless of direct commercial return, a stability argument in its favor. The hybrid architecture itself is a hedge worth watching: state-space layers scale differently with sequence length than attention does, and if the 1M-context era keeps expanding, Mamba-style efficiency may become the next competitive axis.

Ai2's OLMo 3 occupies a different and stricter category: the only frontier-adjacent model that is fully open source rather than merely open-weight. Released November 20, 2025 and updated to OLMo 3.1 on December 12, the 7B and 32B models publish their training data, training code, all checkpoints, and the OlmoTrace system for tracing model outputs back to training data - Ai2. Capability is respectable rather than frontier: OLMo 3-Think 32B scores 96.1 on MATH, edging Qwen 3 32B's 95.4 in its size class.

The distinction OLMo embodies gets flattened in most coverage, and it matters commercially. Every other model on this list gives you weights but not the recipe: you can run DeepSeek V4, but you cannot audit what it was trained on, reproduce it, or trace an output back to its training source. For regulated industries facing provenance requirements, for researchers studying training dynamics, and for anyone whose lawyers ask "prove what is in this model," auditable provenance is not a nice-to-have, it is the product, and OLMo is the only name on the menu. When the EU AI Act's transparency provisions and similar regimes tighten, expect this currently-niche category to become a procurement checkbox that only Ai2 can tick at useful capability levels.

Together, Nemotron and OLMo define the American open-weight position: one model funded by hardware strategy, one by research philanthropy, and, conspicuously, zero funded by the company that used to own this entire category. That absence is the next section.

13. Meta's Exit: How the Llama Era Ended

The previous edition of this guide ranked Llama 3 70B second and described Meta as open-source AI's champion. That framing is now exactly inverted, and the reversal is the most consequential strategic event in open AI since the original DeepSeek moment. The sequence: Llama 4 (Scout and Maverick) shipped in April 2025 to a mixed reception, no Llama 5 followed, and on April 8, 2026 Meta Superintelligence Labs launched Muse Spark, the company's first post-Llama flagship, as a closed model, available through an app and a private API preview only - Meta AI.

Muse Spark itself is technically impressive: natively multimodal reasoning with visual chain of thought, a Contemplating mode that runs parallel multi-agent reasoning, and a 58% score on Humanity's Last Exam, achieved while matching Llama 4 Maverick with over an order of magnitude less compute. We published a full analysis in our Muse Spark guide. But the technology is not the story. The story is that the company whose open releases legitimized this entire category looked at the 2026 landscape and concluded open weights no longer served its interests.

Reasoned from first principles, Meta's logic is legible. Open-sourcing Llama made sense as a commoditize-your-complement strategy when Meta trailed OpenAI and Google: free frontier-adjacent weights devalued rivals' API businesses while costing Meta little, since Meta sells ads, not tokens. By 2026 the strategy had succeeded so completely that it stopped working, because the commoditization was now performed better and faster by DeepSeek, Z.ai, Moonshot, and Alibaba. Meta was paying frontier training costs to give away a product Chinese labs gave away with higher benchmark scores. When your open release is no longer the Schelling point of the ecosystem, the strategic rationale for it collapses.

The practical fallout for teams standardized on Llama is concrete and worth itemizing. Existing weights are not going anywhere: Llama 4 Scout still holds the longest context window of any open model at 10M tokens - Meta AI. But a model line with no successor is a slowly wasting asset in a field moving this fast, and the Llama community license was always more restrictive than the Apache 2.0 and MIT terms now standard among its replacements. Migration paths are friendlier than they have ever been: Qwen3.5 offers a same-size-ladder swap for teams that valued Llama's size range, and DeepSeek V4 Flash offers a cheaper-and-smarter upgrade for most Llama-70B-class API workloads.

The deeper lesson generalizes past Meta and should recalibrate how you read every lab's promises: open weights are a strategy, not an identity. Meta opened when open served it and closed when it did not. Any lab on this list could run the same calculation, which is exactly why section 20's portability habits are not paranoia but basic engineering hygiene. Plan for the ecosystem you can verify today, and notice who now carries the flag Meta dropped.

14. China Now IS Open-Source AI

State the July 2026 leaderboard plainly: every model in the open-weights top six comes from a Chinese lab. GLM-5.2 (Z.ai), MiniMax M3, DeepSeek V4 Pro, Kimi K2.6 (Moonshot), MiMo-V2.5-Pro (Xiaomi), and DeepSeek V4 Flash, per the Artificial Analysis open-weights Intelligence Index - Artificial Analysis. Add Alibaba's Qwen just outside that six and the picture is complete: the open-model frontier lives in Beijing, Hangzhou, and Shanghai. The US counter-lineup, gpt-oss, Nemotron 3 Ultra, Gemma 4, and OLMo 3, competes at the edges (single-GPU, on-device, fully-auditable) rather than at the top.

Why did this happen? First principles again, because the consensus explanation ("China copies fast") explains nothing about why the copying direction reversed. Three structural forces did the work. Export controls made efficiency the objective function: constrained on top-end accelerators, Chinese labs were forced into sparse MoE designs, FP8 and FP4 training, and attention compression, exactly the innovations (49B active of 1.6T, 7x KV-cache cuts) that now define the frontier's economics. Open weights are their distribution strategy: locked out of Western enterprise sales channels, Chinese labs use free weights the way Meta once did, to become the default substrate developers build on. And domestic competition is brutal: at least six Chinese labs ship frontier-adjacent models, and the January 2026 Hong Kong IPOs of MiniMax and Zhipu added public-market fuel to the cadence.

For enterprise buyers, the geopolitical dimension resolves into two separable questions that get wrongly merged. Question one: is it safe to run Chinese-lab weights on your own infrastructure? The weights are static artifacts; they phone home to nobody, they can be audited, red-teamed, fine-tuned, and air-gapped, and MIT or Apache 2.0 terms are enforceable in Western courts. Most security teams that study the question conclude self-hosted open weights are lower-risk than any third-party API of any nationality. Question two: is it safe to send your data to a Chinese-operated API? That is a genuinely different risk profile involving data residency, jurisdiction, and compliance regimes, and many Western enterprises answer no.

The beauty of open weights is that the two questions decouple: you can take the intelligence and leave the API. DeepSeek V4, GLM-5.2, and Kimi all serve through US and EU providers, of which OpenRouter lists dozens with per-provider data policies - OpenRouter. The models are Chinese; your deployment does not have to be. There is also a fragility argument worth keeping on the radar: a top-six built entirely in one jurisdiction concentrates regulatory risk, and an export-policy shift in either direction could reshape release practices faster than any benchmark cycle. Which raises the obvious next question: if open models are this good, how far behind the closed frontier are they really?

15. The Frontier Gap, Quantified

Most open-versus-closed commentary is vibes. The July 2026 data supports actual numbers, so here they are. At the top of the Artificial Analysis Intelligence Index, Claude Opus 4.8 scores roughly 55.7 and GPT-5.5 roughly 54.8, against GLM-5.2's 51 as the open-weights peak - Artificial Analysis. That is about a 9% relative deficit at the very top. One tier down, the open 44-club (DeepSeek V4 Pro, Kimi K2.6, MiniMax M3) sits where the closed frontier sat roughly one product generation ago. Our deep dives on Claude Opus 4.8 and GPT-5.5 cover the closed side of this comparison in detail.

The more important finding is the gap's stability. OpenRouter's June 2026 analysis concludes that open-weight labs have held a consistent 3-6 month capability lag behind the closed frontier for more than 18 months, without the gap widening - OpenRouter. That is a falsifiable thesis, and it keeps failing to be falsified: every closed breakthrough (long-horizon agency, native multimodality, 1M context) has appeared in open weights within two quarters. If the gap were driven by secret architectural moats, it should widen with each closed release. It does not, which suggests the binding constraints are compute and data pipelines, both of which diffuse.

What does a 3-6 month lag mean commercially? It means the question "open or closed?" decomposes by workload horizon. For capability-frontier work, tasks that only the best model on Earth can do at all, closed wins and is worth its premium; a task that is impossible at intelligence 51 and possible at 55.7 has infinite ROI on the difference. For everything already within open-model capability, which after 18 months of gap-closing is the overwhelming majority of enterprise workloads, paying the closed premium is paying for headroom you do not use. The premium is not small: the same OpenRouter analysis puts discounted DeepSeek V4 Flash output around 150x cheaper than GPT-5.5 output.

There is a second-order effect that the stable gap produces: closed pricing power decays on a schedule. Any capability a closed lab charges a premium for today will have an open equivalent within roughly two quarters, at which point the premium must migrate to whatever is newly exclusive. This is visible in how closed labs now market: emphasis has shifted from raw benchmark supremacy toward integration depth, reliability engineering, safety certification, and enterprise tooling, the attributes that do not diffuse with weights. Buyers can exploit the schedule deliberately: adopt closed for the frontier edge, plan the open migration for each workload the moment it stabilizes.

The strategic read for 2026-2027 follows directly. Closed labs must keep finding genuinely new capabilities to justify their tier, because every demonstrated capability gets an open twin and then a 10-100x price collapse. Open labs, meanwhile, compete with each other on the efficiency frontier, which is where all the pricing innovation in section 17 comes from. The gap is real, stable, narrow, and, for most buyers, no longer the deciding variable. The license, increasingly, is.

16. Licenses Decoded: MIT, Apache 2.0, and Display Clauses

"Open source" in 2026 covers at least four legally distinct arrangements, and conflating them is how companies end up with surprise obligations in due diligence. The distinctions were mostly academic when open models were toys; now that they run revenue-bearing products, the fine print is procurement-critical. This section is the decoder, and the one-sentence summary is: the current generation's licensing is the cleanest in open AI history, cleaner than the Llama era, with two attribution quirks worth knowing.

The Modified MIT category is 2026's novel contribution and deserves precision because it sounds scarier than it is. Kimi's license adds exactly one condition to standard MIT: products exceeding 100 million monthly active users or US$20 million per month in revenue must show "Kimi K2" in the interface - Kimi license. Below those thresholds, it is functionally plain MIT. The clause is a branding play, not a control play: Moonshot wants consumer-scale successes to advertise the model. Unless you are building the next TikTok, it will never bind you, but your counsel should know it exists before an acquirer's counsel finds it.

Why does the licensing generation matter beyond compliance hygiene? Because it removes the last structural excuse for enterprise hesitation. The Llama era trained a generation of lawyers to expect open-model licenses with acceptable-use policies, user caps, and naming restrictions, and many procurement processes still budget weeks for model-license review. The 2026 reality is that the top of the market runs on MIT and Apache 2.0, the same licenses as the web servers and databases those enterprises adopted decades ago, with fewer restrictions than the average JavaScript framework's contributor agreement. The review should now take an afternoon.

For teams selecting on license, the practical hierarchy in 2026 runs MIT first, Apache 2.0 a hair behind (its explicit patent grant is actually a plus for enterprises), Modified MIT third with the threshold noted, and custom licenses last, pending legal review. Two additional diligence habits pay off. Check the license of the specific checkpoint, not the family name: labs occasionally ship a flagship under one license and distills under another. And check what the license does not cover: none of these licenses gives you training data or any warranty about its provenance, which is the gap OLMo 3's full openness exists to fill, as covered in section 12. License clarity settles whether you may ship; the next section settles what shipping costs.

17. Pricing Economics: Cache Hits and Cost Per Task

The previous edition of this guide described open models as "5-10x cheaper" than closed ones. That vague multiplier is now both outdated and understated, and 2026's pricing structures deserve real math. The headline mechanism to understand is prefix cache pricing: providers charge a deep discount when your request's input prefix was recently processed, because the expensive prefill computation is already done. DeepSeek turned this into a weapon: V4-Flash charges $0.14 per million input tokens on a cache miss and $0.0028 on a cache hit, a 50x discount, with output at $0.28 - DeepSeek pricing.

Why cache pricing changes agent economics specifically: an autonomous agent re-reads its context on every step. A 20-step agent session over a 200K-token codebase processes roughly 4M input tokens, but 95%+ of them are repeated prefix. At flat frontier pricing that session's input alone costs dollars; at DeepSeek cache-hit pricing it costs about $0.01. This is why MiniMax can market M2.5-Lightning as "$1 to run the model continuously for an hour" - MiniMax. Continuous operation, the defining cost profile of agents, is precisely where the open-model discount compounds from 10x toward the 150x figure OpenRouter measured against GPT-5.5 output on discounted V4-Flash serving - OpenRouter.

Cost per task, not cost per token, is the metric that should drive decisions, and the open labs now publish it directly: Z.ai markets GLM-5.2 coding-agent sessions at roughly $0.26 per task. To make the comparison concrete, take a standard workload: an agent that triages a bug, reproduces it, patches it, and opens a pull request, roughly a SWE-bench-Verified-class task. On GLM-5.2 that is a quarter. On a closed frontier model billed at list prices, the same session commonly lands between $2 and $8 depending on context size and retries. At one task a day the difference is coffee money; at ten thousand tasks a month it is a $20,000-to-$78,000 monthly delta, which is a headcount.

One honest complication keeps the closed premium rational in specific cases: failure costs more than tokens. If the cheaper model succeeds 76% of the time and the expensive one 83%, the 7-point gap means more retries, more human review, and occasionally more damage. Cost per successful task is the true metric, and it sometimes favors the pricier model on hard workloads: a $0.26 task that fails twice before succeeding costs $0.78 plus the latency, and a task that fails silently costs whatever the failure costs. The full methodology for that calculation, including when to route which workload where, is in our LLM cost-efficiency guide. What tokens cost is half the economics; how many tokens fit in one pass is the other half.

18. Context Windows: 1M Tokens Is the New Baseline

In December 2025, this guide treated DeepSeek V3.2's 128K context window as a headline feature. Seven months later, 128K is the floor and 1M is the standard: DeepSeek V4 ships 1M context by default across all official services, GLM-5.2 is 1M, MiniMax M3 is 1M, and MiMo-V2.5-Pro reaches 1M - DeepSeek API Docs. Above them all, the legacy Llama 4 Scout still holds the 10M-token record, a reminder that Meta exited holding the crown for raw length even as its line ended. Output limits scaled too: DeepSeek V4 supports up to 384K output tokens, enough to write a small book in one response.

A million tokens is not an incremental convenience; it changes what a "task" is. It holds roughly 700,000 words: an entire mid-size codebase, a full legal data room, three years of a company's meeting notes, or two hours of transcribed video plus every document it references. Below that threshold, engineers build retrieval pipelines to feed models fragments; above it, for a growing class of problems, you simply load everything and let attention do the retrieval. The engineering effort that went into chunking strategies migrates to context management: what to load, what to cache, what to summarize and when.

The fine print separates marketing context from working context, and buyers should know three distinctions. Native versus extended: Qwen3.5's window is 262,144 tokens native, extensible to ~1,010,000 via YaRN scaling - Hugging Face. Extended windows work, but quality degrades faster toward their limits than trained-native windows like DeepSeek V4's. Supported versus sustained: serving a 1M context at usable speed requires the KV-cache and attention innovations this guide keeps meeting, MiniMax's sparse attention with 15x faster long decode and MiMo's 7x KV-cache reduction; without them, long context is technically supported and economically unusable. Long input versus long-horizon competence: holding a million tokens is not the same as reasoning across them for hours, which is what Terminal-Bench-class evals measure and why section 2's benchmark hierarchy leads with them.

There is also a cost interaction that teams discover the expensive way: long context multiplies input token volume, which is precisely why cache pricing (section 17) and long-context serving efficiency arrived together. A 1M-token context read at DeepSeek's cache-miss rate costs about $0.14; re-read fifty times across an agent session at cache-hit rates, the total stays under a cent. The same session on a provider without prefix caching would cost $7 in input alone. Context strategy and pricing strategy are one decision, not two, and the labs engineering both sides of it are the ones winning the volume war.

The practical guidance falls out cleanly. If your workload is genuinely long-context, whole-repo engineering, discovery review, longitudinal records, shortlist the trained-native 1M models: DeepSeek V4, GLM-5.2, MiniMax M3, MiMo-V2.5-Pro. If your documents fit in 256K, which most do, Kimi K2.6 and native-window Qwen3.5 re-enter the running and you should not pay a premium for length you will not use. And if you plan to run any of these yourself rather than through a provider, context length is also a hardware decision, which brings us to the machine room.

19. Self-Hosting Hardware Tiers in July 2026

Self-hosting open models in 2026 sorts into three hardware tiers, and the single most important concept for navigating them is active versus total parameters. Every flagship on this list is a sparse MoE: DeepSeek V4 Pro holds 1.6T parameters but activates only 49B per token; Kimi K2.6 holds 1T and activates 32B; MiniMax M3 activates 23B of 428B. Compute cost per token tracks the active count, which is why trillion-parameter models are affordable to serve. Memory, however, tracks the total count, since all experts must be resident and ready, and that asymmetry defines the tiers.

The device tier covers phones, laptops, and edge hardware. Gemma 4 E2B and E4B are built for it with audio and vision included, and the small end of the Qwen3.5 ladder (0.8B up through mid sizes) plus Ministral 3 at 3B and 8B run comfortably on consumer silicon - Google DeepMind. This tier is for privacy-mandated work, offline operation, and zero-marginal-cost inference at the edge. The single-GPU tier is defined by one model above all: gpt-oss-120b fits on a single 80GB GPU via MXFP4 quantization, which is the entire reason it tops the download charts - Hugging Face. Quantized 27B-35B models (Gemma 4 31B, Qwen3.6-27B) also live here; our TurboQuant compression guide covers how modern quantization trades size against quality.

The cluster tier is where the frontier MoEs live. Even with FP4/FP8 weights, a 400B-1.6T total-parameter model needs hundreds of gigabytes to multiple terabytes of pooled accelerator memory, meaning multi-GPU nodes or multi-node clusters with fast interconnect, plus the serving expertise to shard experts and manage KV-cache at 1M context. The honest economics: unless you have sustained high utilization, strict data-residency mandates, or fine-tuning needs, renting these models from providers beats owning the metal, and the provider market is deep and competitive - OpenRouter. The break-even point for self-hosting a frontier MoE typically requires keeping the cluster busy the majority of every day, a utilization profile that describes model labs and very few enterprises.

A worked example makes the tiers concrete. A mid-size company wanting private AI for internal documents has three sane configurations. Minimal: Gemma 4 12B or Qwen3.5 mid-size on existing workstation GPUs, adequate for retrieval-augmented Q&A and summarization at zero API spend. Serious: one server with an 80GB accelerator running gpt-oss-120b, a genuine general-purpose intelligence tier for the price of a mid-range car. Frontier: rent DeepSeek V4 or GLM-5.2 from a US or EU provider with a contractual no-retention policy, and reserve the self-hosted cluster conversation for the day utilization data justifies it. Most organizations discover tier one and two cover more of their workload than they expected.

The strategic insight that first-principles buyers act on: because active-parameter counts cluster between 13B and 55B across the entire frontier, serving costs across labs converge, and the real differentiators become memory footprint tricks (MiMo's KV-cache work), cache pricing (DeepSeek), and license terms. Hardware, in other words, has stopped being where open models compete and become where they standardize. What has not standardized at all is how long any given model remains supported.

20. Deprecation Watch and Migration Guidance

The uncomfortable operational truth of 2026 open AI: model lifecycles are now shorter than most companies' budget cycles. Two live examples anchor the point. DeepSeek retires the deepseek-chat and deepseek-reasoner names on July 24, 2026, migrating all traffic to the V4 generation - DeepSeek API Docs. Xiaomi shut down its commercial MiMo-V2 series on June 30, 2026, months after launch, in favor of V2.5 - Xiaomi MiMo. Meta ended an entire model dynasty in one announcement. Deprecation churn is measured in months, and pretending otherwise is how systems quietly break on a Thursday.

Open weights soften this problem in a way closed APIs cannot, and the distinction is worth internalizing. When a closed provider deprecates a model, it is gone; when an open lab moves on, the weights remain downloadable forever, and any provider or your own cluster can keep serving them indefinitely. Deprecation risk in open AI is therefore not availability risk but stagnation risk: the frozen model stops improving while the field moves at the cadence section 1 documented. You will never be forced off DeepSeek V4, but in 2027 you may be competitively unable to stay on it.

A handful of engineering habits convert version whiplash from crisis to routine, and they are cheap to adopt early.

• Abstract the model name: one config value, never hardcoded identifiers scattered through code
• Maintain a golden-task eval set: 50-200 real tasks from your product, runnable against any candidate in hours
• Pin weights, not just names: for self-hosted models, record the exact checkpoint hash you validated
• Subscribe to lab release channels: every lab announces retirements with dates, but only where its users actually look

The deeper pattern behind these habits is treating model choice as a portfolio decision rather than a marriage. The evaluation set is the real asset; models are interchangeable components measured against it. Teams that built golden-task evals in 2025 migrated from Llama to Qwen or DeepSeek in days when Meta exited; teams that validated by vibes are still running two-generation-old models because nobody can prove the replacement is safe. The eval set also converts the field's velocity from threat to option value: every quarterly release becomes a free candidate upgrade you can accept or decline with evidence.

Budget one more thing that almost everyone forgets: prompt and scaffold migration. Prompts tuned for one model's quirks, its verbosity, its tool-call formatting, its refusal patterns, degrade silently on a successor even within the same family. The July 24 DeepSeek migration is mechanical for teams that abstracted their model layer; it is a two-week firefight for teams with three years of prompt folklore hardcoded against deepseek-chat. The 3-6 month capability cadence is not slowing down, so the winning posture makes switching cheap. And switching cheaply matters most in the place where these models increasingly live: inside autonomous agents.

21. From Models to Agents: Putting Open Weights to Work

Everything in this guide converges on one destination, because the destination is where the tokens actually get spent: agents. The benchmark suite went agentic (section 2), the pricing innovations target continuous operation (section 17), and the flagship features, 1M context, parallel sub-agents, native tool use, exist because models now do multi-hour jobs rather than single-turn answers. Kimi K2.5's Agent Swarm coordinating up to 100 parallel sub-agents is not a chat feature - Kimi K2.5 GitHub. Open weights matter more in this world, not less: agents are high-volume, always-on token consumers, exactly the load profile where a 10-150x price advantage compounds daily.

The build-versus-orchestrate decision is the first fork for any team here. Building directly on open weights, choosing a model from this list, wiring up tools, memory, and guardrails, offers maximum control and the economics this guide has priced out; our insider guide to building AI agents walks the full stack, and our primer on making LLMs autonomous covers the conceptual foundations. The honest cost is engineering time: model selection recurs (section 20), serving is a discipline, and multi-agent coordination is a genuinely hard distributed-systems problem, one we examined in our multi-agent orchestration deep dive.

The alternative is running on a platform that has already made these choices and keeps remaking them as the leaderboard churns. O-mega takes that position: an AI workforce platform where agents handle real business processes, research, operations, content, browser-driven tasks, while the platform absorbs the model-selection treadmill this guide documents, routing work to current-generation intelligence without the user re-platforming every quarter. For individual builders who want the self-hosted path instead, our guide to open source personal AI starts from a single machine and one of the device-tier models in section 19.

Model choice interacts with agent design more than most teams expect, and the interaction runs through this guide's rankings. A swarm-style architecture with many parallel workers wants a cheap, fast model per worker, which favors DeepSeek V4 Flash or MiniMax M2.5 economics. A single deep agent grinding through a repository for an hour wants maximum per-step intelligence and cache-friendly context reuse, which favors GLM-5.2 or V4 Pro. A screen-operating agent wants native vision, which shortlists Kimi K2.6 and MiniMax M3. The "best model" question dissolves into "best model per role," and multi-model routing stops being an optimization and becomes the default architecture.

The first-principles way to frame the whole choice: open models turned intelligence into a commodity input, and commodity inputs reward whoever converts them into outcomes most efficiently. For some teams the efficient converter is an in-house agent stack on cache-hit-priced open weights; for others it is a platform that industrializes the conversion. Both answers are correct for different operating profiles, and both get cheaper every quarter the open frontier advances. What neither answer tolerates is the third option most organizations are silently choosing: waiting. The capability is commoditized, the licenses are clean, and the price collapse already happened.

22. Conclusion: Choosing Your Model in July 2026

Seven months rewrote this list from top to bottom, so the conclusion favors decision rules over predictions. The durable facts: the open frontier is a pack of Chinese MoEs within a stable 3-6 months of the closed frontier, licensing is the cleanest it has ever been, 1M context is standard, and pricing innovations like cache-hit billing have collapsed agent operating costs by one to two orders of magnitude. The fragile facts: every specific version number in this guide, which is precisely why section 20's portability habits matter more than any single pick.

The decision path above compresses this guide's ranking into constraints, and it is worth restating in prose. Pick DeepSeek V4 when you want the best overall package of intelligence, price, MIT license, and ecosystem, which is most teams and why it holds #1. Pick GLM-5.2 when peak downloadable intelligence decides the outcome. Pick Kimi K2.6 for multimodal agentic work inside 256K context, MiniMax M3 for long-document economics, Qwen3.5 when you need one architecture from phone to frontier or serious multilingual reach, and gpt-oss-120b or Gemma 4 when the hardware you already own is the constraint. Reserve OLMo 3.1 for provenance-mandated work and Mistral Large 3 for EU-perimeter procurement.

Two habits protect whichever choice you make. Build the golden-task eval set before you commit, because it converts every future model release from a threat into a free upgrade option. And revisit the leaderboard quarterly, not annually: the Artificial Analysis open-weights index and OpenRouter's usage rankings are the two ground-truth instruments this guide trusts, one for capability, one for revealed preference.

This guide was researched and written by Yuma Heymans (@yumahey), founder and CEO of O-mega and co-founder of HeroHunt.ai, who spends an unreasonable share of his week benchmarking open-weight models against the closed frontier to decide what should power production AI agents.

This guide reflects the open-weight model landscape as of July 8, 2026. Model versions, benchmark scores, and pricing in this space change monthly (DeepSeek's legacy model names retire July 24, 2026, sixteen days after this update). Verify current details against the linked primary sources before committing budget.