10 Open-Weight AI Models to Know in 2026

Open-weight models matter because they change who controls the software. You can run one on your own hardware, keep your data in-house, and adjust the model to your task, instead of depending on a provider's API. But the words people use for them are loose, and the licences are not interchangeable. This guide picks ten models from ten developers and records what each model's own card says. It does not rank them.

What does 'open-weight' actually mean?

An open-weight model is one whose trained parameters, the 'weights', are published for download. With the weights you can run the model yourself and, usually, fine-tune it. That is a real freedom, and it is why self-hosting is possible. Our page on open-source and open-weights AI models covers the practical side.

The Open Source Initiative's Open Source AI Definition, version 1.0, sets a higher bar for the label 'open source'. It lists four freedoms: to use the system for any purpose without asking permission, to study how it works and inspect its components, to modify it for any purpose, and to share it with or without modifications. To make those freedoms real it requires three things: sufficiently detailed information about the training data, the complete source code used to train and run the system, and the model parameters, each under OSI-approved terms.

Which ten models, and what do their cards say?

The table is an overview, not a ranking. Rows are in no order of merit. Licence names and sizes are quoted as the model card gives them. 'Date evidence' shows what we could verify: a date on the card itself, a paper the card cites, or the repository creation date from Hugging Face's public API.

What is each model for?

Each description below paraphrases the model's own card. These are the developers' statements of intent, not our findings.

gpt-oss-120b (OpenAI)

OpenAI describes gpt-oss-120b as a general-purpose open-weight reasoning model for production use, designed to fit on a single 80GB GPU. Its card lists the Apache 2.0 licence and about 117 billion parameters with 5.1 billion active, which means only a fraction of the weights is used for each token.

Gemma 4 31B (Google DeepMind)

Gemma 4 comes in five sizes, from E2B and E4B through 12B and 26B A4B up to this 31B dense model. The card describes an instruction-tuned model for chat, reasoning, coding and vision tasks, accepting text and image input and producing text. The licence shown is Apache 2.0.

Qwen3.8-27B (Qwen)

The card calls it a compact, deployment-friendly vision-language model that handles image and video understanding alongside text, with a native context length of 262,144 tokens that can be extended to 1,000,000. This one is Apache 2.0. Note that other Qwen repositories we looked at in Hugging Face's listing, for example Qwen3.8-Flash-Next, are tagged with the licence 'other', so do not assume a family shares one licence.

DeepSeek-V4.1-Flash (DeepSeek-AI)

A multimodal mixture-of-experts model that takes images and text, produces text, and supports contexts up to one million tokens, according to its card. The card states that the repository and the model weights are under the MIT License, which is a permissive licence.

Kimi K3 (Moonshot AI)

Moonshot describes Kimi K3 as an open-weight, multimodal agentic model built for long-horizon coding, knowledge work and reasoning, with 2.8 trillion total and 104 billion active parameters. Both the code and the weights are released under the 'Kimi K3 License', a custom licence. We did not read its terms, so read the LICENSE file before relying on it.

GLM-5.3-Flash (Z.ai)

The card describes a natively multimodal model, the first in the GLM-5 series, meant as a more efficient alternative to GLM-5.2 for coding and agentic tasks, with 320 billion total and 18 billion active parameters. It is listed under the MIT licence. In the same release the full GLM-5.3 repository is tagged 'other', another reminder to check each model individually.

Mistral Small 4 (Mistral AI)

A single model for general chat, reasoning, coding and agentic tasks, with reasoning switched on or off per request, accepting text and image input. It has 119 billion parameters with 6.5 billion activated per token and is licensed under Apache 2.0. The '2603' in the repository name is not explained on the card, so we do not read it as a date.

Phi-4-reasoning-vision-15B (Microsoft)

Microsoft calls it a compact open-weight multimodal reasoning model for vision-language tasks such as solving math and science problems from images and interpreting computer screens, with optional step-by-step reasoning. At 15 billion parameters it is the smallest in this list, and its card states the MIT licence and a March 4, 2026 date.

Olmo 3.1 32B Instruct (Ai2)

The Allen Institute for AI says the model is meant for research and educational use. What sets Olmo apart is how much else the card points to: post-training datasets listed separately, and training and evaluation code in the OLMo-Core, Open-Instruct and OLMo-Eval repositories, under Apache 2.0. The card says only the final few checkpoints are released and that logs are 'coming soon'. That is closer to the Open Source AI Definition than a weights-only release, though we have not assessed whether it meets the definition.

Llama 4 Maverick (Meta)

Meta describes Llama 4 as intended for commercial and research use in multiple languages. Maverick has 17 billion activated and about 400 billion total parameters with 128 experts. The model page is gated behind a login. The licence is the Llama 4 Community License Agreement, which the card calls a custom commercial licence. Visible terms include displaying 'Built with Llama' on related materials and requesting a licence from Meta if a product exceeds 700 million monthly active users.

How should I read the licences?

How do I choose between them without a ranking?

If you are weighing self-hosting against an API, remember that where your data goes is a separate question from which model you pick. Our guide on AI data retention questions lists what to ask. For coding specifically, see models for coding, and for how we handle developer claims, our methodology and the page on vendor claims vs verified findings.

What are the limitations of this list?

• We tested nothing. Every description is the developer's own, taken from its model card.
• Selection is editorial. Ten developers is a sample, not the whole field, and we chose one model per developer for breadth. Omitting a model is not a judgement on it.
• Details change. Licences, sizes and repositories can be edited or replaced. These were read on 18 October 2026.
• Some figures are inconsistent between a card's text and its metadata, such as parameter counts for Qwen, DeepSeek and Llama. We show both where they differ.
• Dates are partial. Several cards state none, so we show the best verifiable evidence and label it.
• The Hugging Face listing data (repository creation dates and licence tags) was retrieved through Hugging Face's public model API and read through an automated summariser, so confirm anything important on the live page.
• Licence terms were not analysed. We name each licence as the card does and summarise only terms visible on the card. This is not legal advice.