Open-Source & Open-Weights AI Models
Open-weights models let you run inference on your own infrastructure, keep data in-house, and fine-tune without depending on a third-party API.
Main model families
Descriptions are deliberately general and not tied to a version or license. Check each model’s card on Hugging Face for its current version, size, and license before you rely on it.
DeepSeek
A Chinese lab that publishes frequent open-weights releases on Hugging Face, including Flash and Pro variants.
Qwen (Alibaba)
A broad family spanning small to very large models.
Kimi (Moonshot AI) and GLM (Z.ai)
Large open-weights releases from two further Chinese labs.
Mistral
A European lab publishing open-weights models alongside commercial ones.
Gemma (Google) and gpt-oss (OpenAI)
Open-weights models from two of the large proprietary labs.
Llama (Meta)
The longest-established family, with wide support across inference tooling.
Why the license matters as much as the benchmark
“Open weights” does not mean “open source” or “free for any use.” Some models use MIT or Apache 2.0, which are permissive. Others add conditions such as revenue thresholds, user-count limits, or acceptable-use policies. Read the license for the exact model and size you plan to deploy.
What self-hosting requires
The largest open-weights models need multi-GPU servers. Smaller and quantized variants run on a single server or a well-equipped workstation, with a trade-off in capability. Budget for hosting, monitoring, and updates, not just the model download.