Open-Source LLMs logopeterindia.net
Knowledge Hub

Open-weight & Open-source Large Language Models

A directory of open-weight and open-source large language models — frontier-scale models, small and efficient families, and the tools used to run them locally.

24 topics•4 categories•Updated November 2026

Frontier-scale Open-weight Models

9 entries

Large models with published weights, with parameter counts and licences from their Hugging Face listings as of 2 November 2026.

01DeepSeek-V4.1-Flash and V4-ProMIT-licensed models from DeepSeek. V4.1-Flash has about 763 billion parameters on its Hugging Face listing and was first published on 10 September 2026, with a 17 October update. V4-Pro-0813 has about 1.65 trillion parameters and was published on 13 August 2026. 02Qwen3.8 familyAlibaba's Qwen3.8 models include Qwen3.8-2.4T-A95B (about 2.4 trillion parameters, custom licence), Qwen3.8-27B (Apache 2.0, August 2026) and Qwen3.8-Flash-Next (about 180 billion parameters, Qwen community licence). 03Kimi-K3Moonshot AI's Kimi-K3 has about 2.78 trillion parameters on its Hugging Face listing, accepts images as well as text, and uses a custom licence. It was first published on 13 June 2026, following Kimi-K2.6 in April 2026. 04GLM-5.3 and GLM-5.3-FlashZ.ai's GLM-5.3 has about 753 billion parameters under a custom licence, and GLM-5.3-Flash has about 321 billion parameters under the MIT licence. Both were published on 25 August 2026. 05MiMo-V2.6Xiaomi's MiMo-V2.6 Pro (about 1.02 trillion parameters) and Flash (about 311 billion parameters) are MIT-licensed reasoning, coding and agent models, published on 21 September 2026. 06Mistral Large 4Mistral's Mistral-Large-4-1T-A52B, named for about one trillion total and 52 billion active parameters, was published on 30 September 2026 with a custom licence. 07GPT-OSS-120BOpenAI's open-weight model under the Apache 2.0 licence, with about 117 billion parameters and 5.1 billion active per token, published in August 2025. It fits on a single 80 GB H100 GPU, and OpenAI has not published a newer open-weight language model since. 08Llama 4 (Scout & Maverick)Meta's April 2025 natively multimodal models, still the newest Llama models in Meta's Hugging Face organisation. Meta launched the proprietary Muse Spark in April 2026 and, according to secondary coverage, released the 30-billion-parameter Muse Glimmer open weights in August 2026, so check Meta's own pages for current licences. 09Gemma 4Google's Gemma 4 open models, from small E2B and E4B variants to 12B and 31B, with the 31B and 12B instruction models under the Apache 2.0 licence. The 31B model was first published in March 2026.

Other Open Model Families

5 entries

Established open and open-weight families.

Run Models Locally

3 entries

Tools for running open models on your own hardware.

Model Libraries & Open Families

2 entries

Where to find and run open-weight and fully open models.

Related Pages

5 entries

Related directories on this site.

Frequently Asked Questions

What does open-weight mean?

The trained model weights are published so you can download, run and often fine-tune the model, though licences differ and training data and code may not be released.

What are the newest open-weight LLMs?

As of 2 November 2026, the Hugging Face listings on this page show DeepSeek-V4.1-Flash (September 2026), Qwen3.8 (August 2026), Kimi-K3 (June 2026), GLM-5.3 (August 2026), MiMo-V2.6 (September 2026), Mistral Large 4 (September 2026) and Gemma 4. New releases arrive often, so check each lab's Hugging Face page.

What happened to Meta's Llama models?

Llama 4, released in April 2025, is the newest Llama in Meta's Hugging Face organisation. Meta launched the proprietary Muse Spark in April 2026 and, according to secondary coverage, released the 30-billion-parameter Muse Glimmer open weights in August 2026. Check Meta's own pages for current models and licences.

Are open-weight models free to use commercially?

Not always. Some models use permissive licences such as MIT or Apache 2.0, while others use custom licences with their own conditions, as the cards on this page note. Read the licence on each model card before commercial use. See the AI Governance Frameworks page on this site for wider obligations.

How is open-weight different from open-source?

Open-source in the strict sense includes licence freedom and open training code and data; many popular models are open-weight with custom licences.