Open-Weight Language Models
Language models whose trained weights anyone can download, from Meta's Llama 2 (2023) to DeepSeek-R1 (2025), can be run and fine-tuned by outsiders.
Open in the interactive tree →Meta released Llama 2 on 18 July 2023 for commercial use under its own licence, Mistral 7B followed on 27 September 2023 under Apache 2.0, DeepSeek-R1 in January 2025 under the MIT licence and OpenAI's gpt-oss on 5 August 2025 under Apache 2.0. Releases differ in how open the licence is, and the Open Source Initiative does not count Llama 2 as open source.
As of October 2026
By early 2026, Alibaba's Qwen had overtaken Meta's Llama in cumulative Hugging Face downloads, and Qwen-derived models made up more than 40% of new language-model derivatives there by August 2025 (Llama about 15%). Moonshot's Kimi K2.5, released in early February 2026, came close to Anthropic's Claude Opus on some early benchmarks at roughly one-seventh of the price; Z.ai and MiniMax listed in Hong Kong in January 2026.
Where AI could help
Medium AI leverage. Models help distil big models into small open ones and write test suites, but frontier training still needs huge compute that no model replaces.
- Distil large models into small ones that run on one machine
- Generate synthetic training data and evaluation sets
- Automate safety testing before weights are released
Prerequisites
Unlocks
- Reasoning Models2024DeepSeek-R1 (January 2025) released an RL-trained reasoning model with open weights
- AI Agents2025Open weights let agent builders run and fine-tune models on their own hardware