The open-weight model landscape in 2026: who publishes what
A plain map of who is publishing downloadable model weights in 2026, under which licences, and what each family is actually good for.
"Open model" now covers a dozen labs with meaningfully different terms. Here is the map, with licences as reported.
The American open releases
- OpenAI gpt-oss — reasoning models in 120B and 20B sizes, under Apache 2.0. Notable mostly because OpenAI publishing weights at all was not a given.
- Meta Llama 4 — multimodal with long-context options, under the Meta Llama Community License. Open weights with conditions, not open source.
- Google Gemma — the Gemma 4 generation is Apache 2.0, sized across a phone-to-cloud range.
- Microsoft Phi — small language models, distributed via Azure, Hugging Face and Ollama.
- NVIDIA Nemotron — publishes weights, training data and recipes, with a 550B model in the family. Unusually complete disclosure.
- Ai2 OLMo — a fully open model flow from the Allen Institute: not just weights, but the whole pipeline. The reference point for genuine research reproducibility.
- Cohere Command A+ — open-weight, enterprise-oriented.
The Chinese open releases
This is where the frontier of open currently sits:
- Kimi K3 (Moonshot AI) — 2.8T parameters, the largest open-weight model reported to date
- GLM-5.2 (Z.ai) — 744B MoE, top of the July 2026 open leaderboards
- DeepSeek V4 — the price leader
- Qwen3.6 (Alibaba) — the licence leader
Europe
- Mistral Large 3 — 675B total / 41B active sparse MoE, 256K context, Apache 2.0
- Magistral Small — Mistral's smaller Apache 2.0 release
How to actually choose
Work down this list in order and you will usually land on the right model in five minutes:
- What can you serve? Your GPU memory sets a hard ceiling. Everything above it is irrelevant, however good.
- What does the licence allow? If you are selling the output, read the licence before the benchmark.
- What is the real cost per request? At your prompt length, with your traffic.
- Does it pass your own evaluation set? Twenty real examples from your business beat any public leaderboard.
Only then does the benchmark table matter.
