Open-Source Model
Also known as: Open Source Model, Open-Source Models, Open-Weight Model, OSS LLM
An open-source model (often open-weight) is a language or multimodal model whose weights are publicly downloadable under a stated license, so you can run, inspect, and further train it yourself. Licenses differ: some are permissive for commercial use; others restrict it. Always read the model card.
These families finished large-scale pretraining (and often instruction tuning) before release. Product work usually starts from one of them, then applies continued pretraining or fine-tuning on your data.
Common families (entry points)
| Family | Org | Links | Notes |
|---|---|---|---|
| Qwen | Alibaba Cloud | qwen.ai, HF, GitHub | Dense and MoE lines; strong multilingual (including Chinese). Common starting point for domain CPT and fine-tuning. |
| gpt-oss | OpenAI | openai.com/open-models, HF, GitHub | Apache 2.0 MoE reasoning models (20B / 120B). Not ChatGPT API weights. Expect OpenAI’s harmony chat format; strong on reasoning and tool use. |
| Llama | Meta | llama.com, HF | Ecosystem default: most tutorials, serving stacks, and fine-tune recipes assume Llama-shaped models. Community license; check commercial terms per release. |
| Mistral | Mistral AI | mistral.ai, HF | Efficient open weights (including Mixtral MoE). Many releases are Apache 2.0. Company also sells closed API models under the same brand. |
| Gemma | Docs, HF | Open weights under Google’s Gemma license (not Apache). Competitive small and mid sizes; later versions add multimodal. | |
| DeepSeek | DeepSeek | deepseek.com, HF | Often MoE; strong on math, code, and reasoning benchmarks relative to size. Popular when you want open weights close to frontier open scores. |
| Phi | Microsoft | HF | Small models trained for quality at low parameter count. Good for local / edge boxes where Llama-70B-class hardware is not an option. |
| OLMo | Allen Institute for AI | allenai.org/olmo, HF | Research-first openness: weights plus training data and code. Usually chosen for reproducibility, not as the strongest chat default. |