Models
Open-weight models: who leads, and how far behind closed models
Kimi, DeepSeek, GLM, Qwen, Gemma, gpt-oss: who leads among open-weight models, the gap with closed ones, the licences and what it means for developers.
On October 27, Artificial Analysis changed the rules. Its Intelligence Index moved to version 5, with 60% private tests, and the gap between the best closed model and the best open one widened from 3 to 12 points (Artificial Analysis). For models whose weights can actually be downloaded, it reaches 26 points. It is a good moment to take stock: who leads, by how much, at what price and under which licence.
A word on vocabulary first. Almost none of these models is "open source" in the free-software sense: they are open-weight models. You can download the trained model, run it yourself and adapt it, but the licence often sets conditions. Unless stated otherwise, scores announced by a vendor are the vendor's own, and prices are in US dollars per million tokens.
Who leads among open models
Chinese labs hold the top of the table.
- Moonshot AI (Kimi). Kimi K3 has 2.8 trillion parameters and a one-million-token context window; Moonshot calls it "the world's first open 3T-class model" (Kimi). Its successor, Kimi K3.1 Pro, has been available through the API since October 23, but its weights will only be published "by Nov 4" (Kimi).
- DeepSeek. DeepSeek V4.1-Flash is a 552-billion-parameter multimodal model with a one-million-token context window (Hugging Face). Its October 17 update, V4.1-Flash-1017, is MIT-licensed (Hugging Face).
- Z.ai. GLM-5.3 weighs about 753 billion parameters according to Hugging Face, and its light version, GLM-5.3-Flash, is MIT-licensed (Hugging Face).
- Alibaba (Qwen). Qwen3.8-27B, a 27-billion-parameter model under Apache 2.0, is one of the most downloaded open models (Hugging Face). The large model, Qwen3.8-2.4T-A95B, is the downloadable version of Qwen3.8-Max, under its own licence (Hugging Face).
On the US side, Meta released Muse Spark Open on October 15: 816 billion parameters, 42 billion of them active per token, and a one-million-token context window. Meta calls it "the most capable US open-weight model to date" (Meta). Google has Gemma 4, whose 31B version is under Apache 2.0 (Hugging Face). OpenAI has released nothing new: its open models page still lists only gpt-oss and gpt-oss-safeguard (OpenAI), and gpt-oss dates from August 5, 2025 (OpenAI).
In Europe, Mistral publishes Mistral Small 4 under Apache 2.0 (Hugging Face) and Mistral Medium 3.5 under a modified MIT licence (Mistral). Its flagship, Mistral Large 4, is listed as proprietary by LMArena (LMArena).
Small models are moving too. Liquid AI released LFM3 on October 27, including a 1.7-billion-parameter version (Liquid AI). Nanbeige4.3-33B, released on October 24 under Apache 2.0, reports 23.8% on Terminal-Bench 4.0 (Hugging Face).
The gap with closed models
On Artificial Analysis's v5 index, the top five models all come from Anthropic and OpenAI. The best open model, Kimi K3.1 Pro, has not published its weights yet. Among weights actually available, DeepSeek V4.1 Flash 1017 and GLM-5.3 lead at 32 points, behind Claude Haiku 5.5, Anthropic's small model (36).
Artificial Analysis Intelligence Index v5, out of 100
█ closed model
▓ open weights, released
░ weights promised, not yet released
Claude Opus 5.5 58 │█████████████████████████████
GPT-6 Astra 57 │█████████████████████████████
GPT-6.1 Sol 55 │████████████████████████████
Gemini 4 Argon 50 │█████████████████████████
Muse Wonder Full 50 │█████████████████████████
Kimi K3.1 Pro 46 │░░░░░░░░░░░░░░░░░░░░░░░
Claude Haiku 5.5 36 │██████████████████
DeepSeek V4.1 Flash 32 │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
GLM-5.3 32 │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
Qwen3.8-27B 20 │▓▓▓▓▓▓▓▓▓▓Source: Artificial Analysis, Index v5, October 27, 2026; Muse Wonder Full and Qwen3.8-27B scores reported by Cybernews, October 28, 2026. Highest reasoning effort for each model.
The gap is widest on long, unseen tasks. On FrontierCode 2.0, a private test where a maintainer must accept the code change, Claude Opus 5.5 completes 46% of tasks and DeepSeek V4.1 Flash 1017, the best open model, 28%. On Terminal-Bench-Science, scientific research run in a terminal, the best released open model tops out at 13%, and most stay below 10%, against 63% for GPT-6 Astra (Artificial Analysis). There is one exception: Kimi K3.1 Pro has the best score of any model on Terminal-Bench 4.0, at 67%.
The reshuffle has its critics. Several open models lost more than 10 points at once, such as DeepSeek V4.1 Flash, down from 45 to 32, and commenters on Hacker News are asking for explanations (Cybernews).
The LMArena ranking, based on votes from users who compare two blind answers, shows a shorter gap.
| Model | Rank | Score | Licence |
|---|---|---|---|
| Gemini 4 Argon | 2 | 1,520 | proprietary |
| Claude Opus 5.5 | 3 | 1,509 | proprietary |
| Kimi K3 | 19 | 1,488 | Kimi K3 licence |
| GLM-5.3 | 28 | 1,481 | MIT |
| DeepSeek V4.1-Flash | 35 | 1,476 | MIT |
| Gemma 4 31B | 77 | 1,452 | Apache 2.0 |
| Mistral Medium 3.5 | 122 | 1,426 | modified MIT |
| gpt-oss-120b | 219 | 1,352 | Apache 2.0 |
Source: LMArena, text leaderboard of October 27, 2026, 419 models. First place goes to Gemini 4 Carbon (1,528), a score still marked "preliminary".
Our reading: the two rankings do not measure the same thing. In user votes, the best open model, Kimi K3, is only 32 points behind Gemini 4 Argon. On long agent tasks with unseen tests, the gap is much wider. That is precisely the use that matters for a coding agent.
Cost, the first argument for open models
Price remains the big advantage of open models, but not for all of them, and no longer automatically.
Official API price, in $ per million output tokens
█ closed model
▓ open-weight model
Claude Opus 5.5 20 │████████████████████
Kimi K3 15 │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
GPT-6.1 Sol 10 │██████████
Gemini 4 Argon 10 │██████████
Qwen3.8-Max 6 │▓▓▓▓▓▓
GPT-6.1 Terra 5 │█████
GLM-5.3 4.40 │▓▓▓▓
DeepSeek V4-Pro 3.96 │▓▓▓▓
DeepSeek V4.1-Flash 1.20 │▓
Claude Haiku 5.5 0.50 │█
GPT-6 Luna 0.50 │█
GLM-5.3-Flash 0.50 │▓Sources: Anthropic, Kimi, OpenAI, Google (launch price until November 30), Alibaba Cloud (international rate; this is the price of the hosted version), Z.ai, DeepSeek (peak hours, half price off-peak). Claude Haiku 5.5: prompts of 100,000 tokens or fewer. Checked on October 28, 2026.
DeepSeek V4.1-Flash remains the most striking example. It costs $0.30 for input and $1.20 for output at peak hours, and half that off-peak (DeepSeek). By contrast, Kimi K3 is billed at $3 for input and $15 for output, more than GPT-6.1 Sol (Kimi). Small closed models have also reached the bottom of the scale: Claude Haiku 5.5 and GPT-6 Luna cost $0.10 and $0.50 (Anthropic, OpenAI).
Price per token is not the whole story. Artificial Analysis also measures the cost of a full task in its index: $10.04 for Claude Opus 5.5, $6.24 for GPT-6 Astra, $1.27 for GPT-6.1 Sol and $0.96 for GPT-6.1 Terra (Artificial Analysis). A cheap model that writes many reasoning tokens can end up costing as much as a leaner closed model.
Licences: open does not mean free
Licences form a spectrum, from truly permissive to conditions aimed at large competitors.
| Model | Licence | What to know |
|---|---|---|
| gpt-oss-120b and 20b | Apache 2.0 | Commercial use allowed, subject to the gpt-oss usage policy (OpenAI) |
| Gemma 4, Qwen3.8-27B, Mistral Small 4, Nanbeige4.3-33B | Apache 2.0 | Standard software licence (Hugging Face) |
| DeepSeek V4.1-Flash-1017, GLM-5.3-Flash | MIT | Standard software licence (Hugging Face) |
| GLM-5.3 | GLM-5.3 licence | Security review by Z.ai for API providers above $10bn in revenue over 12 months (Hugging Face) |
| Kimi K3 | Kimi K3 licence | Separate agreement for API providers above $20M in revenue over 12 months; "Kimi K3" credit above 100M monthly users or $20M in monthly revenue (Hugging Face) |
| Qwen3.8-2.4T-A95B | Qwen3.8-Max licence | Separate licence for API and assistant services above $50M in revenue over 12 months (Hugging Face) |
| MiniMax M3 | MiniMax Community License | Commercial use must be notified to MiniMax; written authorisation above $20M in annual revenue (Hugging Face) |
| Muse Spark Open | Muse Model License | "Built with Muse Spark Open" credit required; agreement with Meta for API providers above $20M over 12 months; download on request (Hugging Face) |
| LFM3 | LFM Open License v1.0 | Commercial use not covered above $10M in annual revenue (Hugging Face) |
Our reading: the trend is clear. The largest open models come with custom licences whose clauses mostly target hosts that would resell the model as an API. For a team using it internally, most remain usable. For a software vendor above the thresholds, read the licence before choosing.
What actually gets downloaded
The models that make headlines are not the ones developers download most.
Hugging Face downloads last month, in millions
Gemma 4 31B 8.47 │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
Qwen3.8-27B 7.52 │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
GLM-5.3-Flash 6.42 │▓▓▓▓▓▓▓▓▓▓▓▓▓
gpt-oss-20b 6.01 │▓▓▓▓▓▓▓▓▓▓▓▓
gpt-oss-120b 2.83 │▓▓▓▓▓▓
GLM-5.3 2.30 │▓▓▓▓▓
DeepSeek V4.1-Flash 1.85 │▓▓▓▓
Kimi K3 0.85 │▓▓
Muse Glimmer 30B 0.21 │▏
Mistral Small 4 0.07 │▏
Muse Spark Open 0.01 │▏Source: "Downloads last month" counter on each Hugging Face page (Gemma 4 31B, Qwen3.8-27B, GLM-5.3-Flash, gpt-oss-20b, gpt-oss-120b, GLM-5.3, DeepSeek V4.1-Flash, Kimi K3, Muse Glimmer 30B, Mistral Small 4, Muse Spark Open), checked on October 28, 2026. Main repository of each model, excluding quantised versions and community copies.
Two caveats. Hugging Face counts every request to certain files, such as config.json, including those from automated pipelines (Hugging Face). And Muse Spark Open is only available on request, which mechanically holds back its downloads. The ranking still shows a clear trend: models of 20 to 30 billion parameters, which fit on a single machine, dominate.
What it means for developers
Locally, hardware decides before scores do. gpt-oss-20b runs with 16 GB of memory and gpt-oss-120b on a single 80 GB GPU (OpenAI). Muse Spark Open needs two Mac Studio M5 Ultra machines (Meta). For Kimi K3, Moonshot recommends clusters of 64 accelerators or more (Kimi). Our guide to running a coding model locally with Ollama walks through the setup.
In coding agents, an open model now plugs in like a closed one. DeepSeek's API accepts the OpenAI and Anthropic formats, and its documentation explains how to use it with Claude Code, Codex and OpenCode (DeepSeek). Switching models often comes down to an API address and a key.
In companies, adoption remains low. According to Menlo Ventures, the share of open-source models in enterprise API usage fell from 19% in 2024 to 11% in 2025, and Chinese models account for only about 1% of that usage (Menlo Ventures). Menlo Ventures is an Anthropic investor.
The next test has a date: Moonshot promises Kimi K3.1 Pro's weights by November 4. If it delivers, the best downloadable model would be 12 points from the top of the Artificial Analysis index, instead of 26 today.
Sources
- Artificial Analysis, Intelligence Index v5, October 27, 2026, checked on 2026-10-28
- Cybernews, open-weight models reshuffle, October 28, 2026
- LMArena, text leaderboard, October 27, 2026, checked on 2026-10-28
- Kimi K3 and Kimi K3.1 Pro, Moonshot AI
- Meta, Muse Spark Open, October 15, 2026
- OpenAI, gpt-oss, August 5, 2025, and open models page
- DeepSeek, API models and pricing, checked on 2026-10-28
- Z.ai, pricing, checked on 2026-10-28
- Alibaba Cloud, Model Studio pricing, checked on 2026-10-28
- Anthropic, pricing and OpenAI, API pricing, checked on 2026-10-28
- Liquid AI, LFM3, October 27, 2026
- Hugging Face, how downloads are counted
- Menlo Ventures, State of Generative AI in the Enterprise 2025, December 9, 2025
Spotted an error in this article? Report it · Our corrections policy
