// the take
A single markdown table tracking open-weight LLMs that carry licenses permissive enough for commercial use, plus separate tables for code models, pretraining datasets, and instruction/alignment-tuning datasets. Useful for anyone picking a base model for a product and needing to know upfront whether the license actually allows that, not just whether weights are public.
The license column does real work other lists skip — it distinguishes genuinely permissive licenses (Apache 2.0, MIT) from the fine-print-restricted ones (Llama's 700M-user cap, Qwen's ban on using outputs to train competing LLMs, DeepSeek's derivative clauses). Each entry links out to checkpoints, the paper/blog, and context length, so it functions as a real reference rather than just names. Splitting general LLMs, code LLMs, and dataset tables keeps the scope honest instead of dumping everything into one undifferentiated list.
It's a hand-maintained markdown file with no JSON/API export, so using it programmatically means scraping a markdown table. The field is moving fast and the newest rows here are from early 2025 — no Llama 3.3, Qwen2.5/3, or DeepSeek-V3, so anything shipped in the last year is absent and you can't tell from the repo whether that's a backlog or abandonment. The license summaries are informal paraphrases, not legal review; relying on 'Apache 2.0' or 'free with usage restriction' here instead of reading the actual license text is exactly how someone ends up out of compliance. There's no stated inclusion bar beyond 'commercial-use licensed,' so there's no way to judge whether a model's absence means it was missed or deliberately excluded.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →