// The State of Open Models 2026 · about six minutes

Executive summary

The most downloaded models are not necessarily the most used. And what gets used does not necessarily take the money.

Illustration: layered landscape with a pale sun

Open models have moved to the front page of the conversation. They are downloaded hundreds of millions of times, take a large share of the tokens on inference platforms, and on one programming benchmark some of them solve a task for a fraction of what closed models cost. Counted four different ways, they give four different leaders.

Part one

What is built and how good it is

Who makes the open models people download and how far the best of them are from the best models evaluated.

01 Ecosystem · Chapter 01

Each source has a different leader

From October 2025 to September 2026 Qwen has taken 54% of the downloads of models on Hugging Face, and eight of the ten most downloaded are from the Qwen family. Its most downloaded model, Qwen3-0.6B, was downloaded twice as often as the most downloaded model of Meta, OpenAI or Google. The size of this family’s models has a lot to do with it.

The most downloaded model of each of the eight largest labs, October 2025 to September 2026

Qwen Other labs

  1. QwenQwen3-0.6B

    202M downloads

    54%

  2. MetaLlama-3.1-8B-Instruct

    97M downloads

    8%

  3. OpenAIgpt-oss-20b

    81M downloads

    3%

  4. Googlegemma-4-26B-A4B-it

    62M downloads

    8%

  5. DeepSeekDeepSeek-OCR

    34M downloads

    6%

  6. Mistral AIMistral-7B-Instruct-v0.3

    29M downloads

    3%

  7. MicrosoftFlorence-2-base

    13M downloads

    2%

  8. NVIDIALlama-3.1-Nemotron-Nano-VL-8B-V1

    11M downloads

    3%

Hugging Face downloads of the tracked open models over twelve monthly readings (1 November 2025 to 1 October 2026). Source: Hugging Face hub-stats; ATOM Project.

Conclusion

Most of its downloads are of small models: seven in ten Qwen downloads are of models under 10 billion parameters. But small is not enough, since others publish small models too, and Qwen also leads among models of 10 billion parameters or more.

How we measured it

Evidence E3

Downloads, not users or deployments.

02 Capability · Chapter 02

One open model is about 2 points from the best in science and agentic work; elsewhere the gap is 17 to 32

On Epoch AI’s own tests, the best open-weight model was 2.2 points behind the best model of all in graduate-level science, 16.8 in advanced mathematics and 31.8 in factual knowledge. In long agentic tasks, run by Vals AI with the same agent for every model, it was 1.5 points behind. The best closed models right now are from Google, OpenAI and Anthropic. Among the open ones, Kimi K3.1 Pro, whose weights were published on 4 November, leads all four.

Best open-weight model against the best closed model

Best open-weight model Best closed model

0%25%50%75%100%Graduate-level scienceKimi K3.1 Pro 94.2%Gemini 4 Argon 96.4%2.2 ptsAgentic workKimi K3.1 Pro 63.6%Claude Fable 5.2 and Claude Opus 5.5 65.2%1.5 ptsAdvanced mathematicsKimi K3.1 Pro 78.2%Claude Fable 5.2 95.1%16.8 ptsFactual knowledgeKimi K3.1 Pro 62.2%Claude Fable 5.2 94.0%31.8 pts
  • Graduate-level science

    Kimi K3.1 Pro 94.2%

    Gemini 4 Argon 96.4%

    2.2 pts

  • Agentic work

    Kimi K3.1 Pro 63.6%

    Claude Fable 5.2 and Claude Opus 5.5 65.2%

    1.5 pts

  • Advanced mathematics

    Kimi K3.1 Pro 78.2%

    Claude Fable 5.2 95.1%

    16.8 pts

  • Factual knowledge

    Kimi K3.1 Pro 62.2%

    Claude Fable 5.2 94.0%

    31.8 pts

Share of each test answered or solved, models released by 10 November 2026; the number under each bar is the distance in percentage points. Three tests in Epoch AI’s own runs; agentic work in Vals AI’s Terminal-Bench 4.0, the same agent for every model. Sources: Epoch AI (CC BY 4.0); Vals AI.

Conclusion

The distance moves with each release: what the best models achieve, an open model reaches 2 to 11 months later. For choosing a model, the task matters more than any ranking.

How we measured it

Evidence E3

Three tests in Epoch AI’s own runs and one, Terminal-Bench 4.0, in Vals AI’s; models released by 10 November 2026.

Part two

What is used and what it costs

Which open models are used most and what it costs to complete a task with each of them.

03 Usage · Chapter 03

A third of the open-weight tokens went to DeepSeek

From October 2025 to September 2026, DeepSeek’s models took 31% of the open-weight tokens on OpenRouter; then Tencent, Xiaomi and Z.ai. Qwen, which leads the downloads, took 1.0%. Over the year, open-weight models took 60% of the tokens in OpenRouter’s daily top 50 ranking.

Open-weight token consumption in OpenRouter’s ranking, October 2025 to September 2026

DeepSeek Other developers

  1. 1

    DeepSeek V4 Flash DeepSeek

    18.1% of open tokens

  2. 2

    MiMo-V2.5 Xiaomi

    9.8% of open tokens

  3. 3

    Hy3 Tencent

    7.5% of open tokens

  4. 4

    GLM 5.3 Flash Z.ai

    6.3% of open tokens

  5. 5

    Hy4 preview Tencent

    5.8% of open tokens

  6. 6

    DeepSeek V4.1 Flash DeepSeek

    5.0% of open tokens

  7. 7

    DeepSeek V4 Pro DeepSeek

    4.6% of open tokens

  8. 8

    Nemotron 3 Ultra NVIDIA

    4.6% of open tokens

Denominator: tokens processed by open-weight models in OpenRouter’s daily ranking. Top 50. Source: OpenRouter (openrouter.ai/rankings), as of 15 October 2026 (CC BY 4.0).

Conclusion

The models in use change within months: no single open model led for more than 7 months, while DeepSeek was the most used open lab in 19 of 21 months.

In the sample of requests that OpenRouter sorts by subject, a closed model is the most used in all 17 subjects, from software to law. Among the open ones, DeepSeek V4.1 Flash is the most used in 11 of them; the best place an open model reaches is second, GLM-5.2 in relationships and Kimi K3 in health & wellness.

How we measured it

Evidence E3

Traffic on one platform that routes requests to many models: not users, and not market share.

04 Cost · Chapter 04

At the top, an open model finished tasks for less than half the price, but not every open model is cheap

On the same 66 long agentic tasks, run with the same agent by Vals AI, Kimi K3.1 Pro, open-weight, completed 63.6% at $7.18 per solved task; Claude Opus 5.5 completed 65.2% at $18.48 and Claude Fable 5.2 65.2% at $22.35.

Cost per solved task, eight models on the same 66 long agentic tasks

Open weights Closed

  1. 1

    Muse Ember 39.4% completed · 26 min a task

    $0.55 per solved task

  2. 2

    GPT-6.1 Sol 55.0% completed · 45 min a task

    $3.13 per solved task

  3. 3

    GLM-5.3 Flash 37.4% completed · 61 min a task

    $3.26 per solved task

  4. 4

    Kimi K3.1 Pro 63.6% completed · 49 min a task

    $7.18 per solved task

  5. 5

    Claude Sonnet 5.5 64.1% completed · 82 min a task

    $16.76 per solved task

  6. 6

    Claude Opus 5.5 65.2% completed · 64 min a task

    $18.48 per solved task

  7. 7

    Claude Fable 5.2 65.2% completed · 55 min a task

    $22.35 per solved task

  8. 8

    GLM-5.3 38.9% completed · 83 min a task

    $24.11 per solved task

Terminal-Bench 4.0 with the same agent (mini-SWE-agent), three runs, at the serving provider’s API prices. Source: Vals AI.

Conclusion

Open weights do not make a task cheap: GLM-5.3 cost $24.11 per solved task, more than most closed models, and GPT-6.1 Sol, closed, $3.13. What matters is how many tokens a model spends, at what price, to finish. The saving also has a price in time: the median open model took 70 minutes per task, the median closed one 54.

How we measured it

Evidence E3

One benchmark of long agentic tasks, with one agent; costs at the serving provider’s prices in each run. Models released by 10 November 2026.

Part three

Who pays and where they run

How much of the money goes to open models, who uses them, and on whose infrastructure.

05 Adoption · Chapter 05

In US enterprises, by Menlo Ventures’ estimate, open models took a fifth of the tokens and 4% of the money

In Menlo Ventures’ modelled estimate for US enterprises, open models processed 19% of the tokens in September 2026 and took 4% of the spending on AI models. Counted by companies that pay, the figure is smaller: 6.4% of the US businesses on Ramp that pay for AI pay a platform that serves open models.

Open models’ share of the tokens and of the spending, in the two sources that measure both

Menlo Ventures US enterprises, modelled, Sept. 2026

Tokens19%

Spending4%

Vercel AI Gateway developers and startups, Sept. 2026

Tokens71.5%

Spending23%

Menlo Ventures: a modelled estimate built on a survey of about 400 US decision-makers (Morning Consult, September 2026). Vercel: tokens through its AI Gateway and spending estimated at list prices. Figures as published.

Conclusion

Where they are already in use, they grow fast: on Vercel’s gateway, used by developers and startups, open models went from 13% of the tokens in April to 71.5% in September, for 23% of the spending. In one panel of 200 large companies, 42% had them in production, up from 31% in July; almost every business that pays a platform serving open models also pays OpenAI or Anthropic.

How we measured it

Evidence E2

Published figures of different sources, each with its own population; none is a count of companies in general.

06 Deployment · Chapter 06

If the weights are open, where does the model run?

Independent hosts served 39% of the tokens routed through OpenRouter in the month to 9 November 2026, open and closed models together; the model makers took 52% and the big clouds 9%. On their own price lists, many hosts charge what the developer charges.

Who served the tokens routed through OpenRouter, a month to 9 November 2026

Independent hosts Model makers Big clouds

39%52%9%

When does renting your own GPUs pay off? 8x B200 rented with DeepSeek-V4-Pro cost the same as DeepSeek’s API if they work this share of the month

The shorter the bar, the sooner renting pays off; “never” means the API is always cheaper.

  • The scenario in chapter 0656% of the month
  • If the GPU hour costs the median spot price33% of the month
  • If the GPU hour costs the highest list price96% of the month
  • If each request carries 10 input tokens per output token26% of the month
  • If each request carries 1 input token per output token84% of the month
  • If the GPUs generate as fast as in a single-turn test77% of the month
  • If you compare with the cheapest API on OpenRouternever
Top: monthly tokens on OpenRouter’s providers page, open and closed models together, grouped by us. Bottom: a modelled scenario of GPU load, to see when self-hosting a model pays off. Sources: OpenRouter; providers’ price pages; SemiAnalysis InferenceX.

Conclusion

Whether running a model yourself is cheaper depends on how busy the GPUs are kept and on four assumptions: what the GPU hour costs, how many input tokens each request carries per output token, how many tokens a second the GPUs actually generate, and which API price you compare against. In the chapter’s illustrative scenario, eight rented B200s, about $48,700 a month, cost the same as DeepSeek’s API for DeepSeek-V4-Pro if they work 56% of the month; changing one assumption at a time moves that from a quarter of the month to never.

How we measured it

Evidence E2

Traffic routed through one platform, as displayed on 9 November 2026; the break-even is a modelled scenario, not an observed cost.

General conclusions of the report

The open-model ecosystem changes and takes new shape every month. But this last year helps us understand some of the questions we want to answer about models, families, use, cost and deployment.

The most surprising

The most downloaded is not the most used. From October 2025 to September 2026 Qwen took 54% of the downloads on Hugging Face and 1.0% of the open-weight tokens in the daily top 50 of OpenRouter, where DeepSeek led with 31%.

What is used most is not what companies pay most for. Open models processed 19% of US enterprises’ tokens and took 4% of the spending, in Menlo Ventures’ modelled estimate. Budgets still go to the big labs of closed models.

Open weights: where do they run? Independent hosts, model makers and the big clouds share the traffic one router sends, many hosts charge what the developer charges, and some of the cheaper offers declare a lower precision.

The lever that moves the shift to open models

These are our interpretations, not measurements. Compare models by what a finished task costs and how long it takes, not by the price of a token, on tasks that resemble your own. On the one benchmark we cost, some open-weight models solved tasks for a fraction of what closed models cost and others cost more, and the closed models were usually faster. For a team that does not yet know its traffic, an API avoids committing to capacity; capacity of one’s own is worth testing for large, steady volumes or for data that must stay in-house.

The open-model market has changed faster than the questions we ask of it. Downloads, tokens, money and how the models are served give us a general idea of what the whole ecosystem looks like.