The State of Open Models 2026
Capability · Usage · Cost · Adoption · Deployment
How the open-model ecosystem looks today, read from the most important sources in the field.

// About this report
What is the current state of open models?
The conversation revolves around whether open models are on a par with proprietary or closed models. But for companies, the decision goes beyond a benchmark. The questions companies ask most are not answered by a comparison of scores. What concerns them is something else: how many companies use them, what it costs them to complete a task, or who operates the infrastructure that runs them. The idea behind this report is simple: before drawing a conclusion about open models, you need to know what the ecosystem looks like.
Each chapter presents its results in charts. Its Evidence & Methods page sets out the sources, calculations and limitations, so that you can check our work.
The report is written by Helmcode Research, part of Helmcode. We are an inference platform and cloud for open-weight models.
00 Executive summary
What did we find? A quick summary
-
Chapter 01 · Ecosystem
Each source has a different leader
54%
of the downloads of the models we track on Hugging Face from October 2025 to September 2026 went to Qwen. 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 Qwen's models, and the fact that they run on home graphics cards, have given them a very wide distribution.
A different leader in each source
Downloads
Hugging Face, Oct. 2025–Sept. 2026
- 1
Qwen
54.2% of the downloads
- 2
Google
8.5% of the downloads
- 3
Meta
8.2% of the downloads
- 4
DeepSeek
6.1% of the downloads
Tokens
OpenRouter, open weights, Oct. 2025–Sept. 2026
- 1
DeepSeek
31.3% of the tokens
- 2
Tencent
16.5% of the tokens
- 3
Xiaomi
12.2% of the tokens
- 4
Z.ai
12.1% of the tokens
Each lab’s share within its source; two populations and two units that are not added and are not market shares. Sources: Hugging Face hub-stats and the ATOM Project; OpenRouter (CC BY 4.0). If most downloads go to a few families, how close do the best open models come to the best?
- 1
-
Chapter 02 · Capability
What open models can do depends on the task
2–32points
of difference between the best open-weight model and the best closed model on four tests: 2.2 in graduate-level science, 16.8 in advanced mathematics, 31.8 in factual knowledge and 1.5 in long agentic tasks, on Epoch AI’s and Vals AI’s tests. Since its release, Kimi K3.1 Pro is the open model closest to the closed ones.
Best open-weight model against the best closed model
Best open-weight model Best closed model
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. Capability is one thing. But which models do people actually use?
-
Chapter 03 · Usage
Over the whole year, a third of the open-weight tokens went to DeepSeek alone
31%
of the open-weight tokens on OpenRouter from October 2025 to September 2026 went to DeepSeek’s models; Tencent (17%), Xiaomi (12%) and Z.ai (12%) followed. Qwen, the most downloaded, took 1.0%. Downloads may point to local use, but not to use through an API.
Open-weight token consumption in OpenRouter’s ranking, October 2025 to September 2026
DeepSeek Other developers
- 1
DeepSeek V4 Flash DeepSeek
18.1% of open tokens
- 2
MiMo-V2.5 Xiaomi
9.8% of open tokens
- 3
Hy3 Tencent
7.5% of open tokens
- 4
GLM 5.3 Flash Z.ai
6.3% of open tokens
- 5
Hy4 preview Tencent
5.8% of open tokens
- 6
DeepSeek V4.1 Flash DeepSeek
5.0% of open tokens
- 7
DeepSeek V4 Pro DeepSeek
4.6% of open tokens
- 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). Tokens measure volume. What does a finished task actually cost?
- 1
-
Chapter 04 · Cost
At almost the same quality, the best open model finished a task for less than half the price
$7per solved task
with Kimi K3.1 Pro, open-weight, which completed 63.6% of the same 66 long agentic tasks; Claude Opus 5.5 completed 65.2% at $18.48. But GLM-5.3 cost $24.11 per solved task and GPT-6.1 Sol, closed, $3.13: a cheap token is not a cheap task.
Vals AI, Terminal-Bench 4.0.
Cost per solved task, eight models on the same 66 long agentic tasks
Open weights Closed
- 1
Muse Ember 39.4% completed · 26 min a task
$0.55 per solved task
- 2
GPT-6.1 Sol 55.0% completed · 45 min a task
$3.13 per solved task
- 3
GLM-5.3 Flash 37.4% completed · 61 min a task
$3.26 per solved task
- 4
Kimi K3.1 Pro 63.6% completed · 49 min a task
$7.18 per solved task
- 5
Claude Sonnet 5.5 64.1% completed · 82 min a task
$16.76 per solved task
- 6
Claude Opus 5.5 65.2% completed · 64 min a task
$18.48 per solved task
- 7
Claude Fable 5.2 65.2% completed · 55 min a task
$22.35 per solved task
- 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. Outside the benchmarks and platforms, how many organisations use an open model?
- 1
-
Chapter 05 · Adoption
In US enterprises, by Menlo Ventures’ estimate, open models took a fifth of the tokens and 4% of the money
4%
of US enterprises’ spending on AI models went to open models in September 2026, while they processed 19% of the tokens (Menlo Ventures). Only 6.4% of US businesses that pay for AI pay a platform serving open models (Ramp).
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. And when they do, whose infrastructure does the model run on?
-
Chapter 06 · Deployment
Published weights do not say where a model runs
39%
of the tokens routed through OpenRouter in a month went to independent hosts such as Together, DeepInfra or Wafer, which serve mostly open-weight models; labs serving their own models took 52% and the big clouds 9%.
Illustration: the node in the chapter’s break-even scenario Who served the tokens routed through OpenRouter, a month to 9 November 2026
Independent hosts Model makers Big clouds
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.
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.
The executive summary has each finding with its chart, which you can read in about six minutes. The Full Report has every chapter, with sources one click from each chart.
// Conclusions
What the data tell us about the state of open models
Five simple conclusions. If you want to understand them better, each one points to the chapters and the sources.
There are different leaders depending on what is observed
Qwen took 54% of the downloads on Hugging Face from October 2025 to September 2026, and nearly half of the derivative models. In the same months DeepSeek had 31% of the open-weight tokens in OpenRouter’s ranking, and Kimi K3.1 Pro had the best open score on the report’s capability tests. Downloads, tokens and tests measure different things: the model most used locally, the most used model overall and the most capable model. None of the three coincide.
Models change fast; labs, less so
In September 2026, 76% of the tokens in OpenRouter’s top 50 went to versions that had entered it in the previous 90 days. DeepSeek, by contrast, was the most used open lab in 19 of 21 months. A version leaving the top places means another gets more traffic, not that it falls out of use.
The distance to closed models depends on the task
Kimi K3.1 Pro came within 2.2 points of the best closed model in graduate-level science and 1.5 in long agentic tasks, but was 16.8 behind in advanced mathematics, 21.7 in scientific images and 31.8 in factual knowledge. Open models are one step away from closed ones in agentic programming, but depending on the task, some labs are closer than others.
The price of a token is not the cost of a task
In US enterprises open models processed 19% of the tokens and received 4% of the spending, according to Menlo Ventures. On Terminal-Bench 4.0, picking the model with the cheaper token is not enough: in 197 of 741 pairs of models, that choice spent more per solved task. In 18 of 28 pairs of an open and a closed model with similar success, the open model was cheaper. Running a model on your own machines does not pay by default either: it depends on how busy the GPUs are kept.
Open weights widen the options
Publishing the weights lets others serve a model, but on OpenRouter the labs themselves served 52% of the tokens in the month to 9 November 2026. And almost every company that pays an open-model platform on Ramp keeps paying OpenAI or Anthropic.
01–06 The Full Report
Six chapters
-
Ecosystem
Each Source Has a Different Leader
Who makes the open models people download, and how quickly that map has changed.
-
Capability
How Close Are Open Models to Closed Ones?
How far the best open models are behind the best, in months and in points.
-
Cost
The Economics of Open Models
What a finished task costs with each model, and how long it takes.
-
Adoption
Who Actually Uses Open Models?
Who uses open models, and how much of the spending goes to them.
-
Deployment
Who Runs Open Models?
Where open models run: providers, APIs, GPUs and the cost of hosting them.
PDF Get the whole report
The complete report
The executive summary and the six chapters with every chart, and a short note on where each chart comes from, in one document.
Method
This edition draws from
650
Days of daily OpenRouter rankings
23,494
Research works naming a model family, 2025
47
Models costed per solved task
17
Public data sources






