PhyseaWiki How AI actually works Papers physea.ai →

Model families

What are Llama, Qwen, DeepSeek, Mistral, and Gemma known for?

Llama, Qwen, DeepSeek, Mistral, and Gemma are the leading open-weight families: models whose files you can download and run yourself. Each has its own reputation, from Llama's role in starting the open-weight wave to Gemma's focus on small, on-device sizes.

Last updated 2026-06-15 · Physea Labs

Open-weight families publish their trained model files for download, so you can run, inspect, and adapt them on your own hardware. Five families carry most of the weight here in 2026, and each has its own character.

Llama, from Meta, is the family that opened the open-weight wave and gave many builders their first downloadable, capable model. Llama 4 uses a mixture-of-experts design and is natively multimodal, and Meta releases it under its own community license rather than a fully permissive one.[1] The license point matters: open weight does not always mean unrestricted use.

Qwen, from Alibaba Cloud’s Qwen team, is known for coming in many sizes and for strong multilingual coverage. Qwen3 ships in a wide range of sizes and supports more than 100 languages, with its open-weight models released under the permissive Apache 2.0 license.[2] That spread of sizes makes it a common starting point when you need a model to fit a specific budget or device.

DeepSeek, from DeepSeek-AI, is known for cost-efficient open reasoning. DeepSeek-V3 is a large mixture-of-experts model (671B total parameters with 37B active per token) released with downloadable weights, and its model card reports an unusually efficient training run measured in GPU hours.[3] The code is MIT-licensed, one of the more permissive choices, and the family drew wide notice for strong reasoning from an open release.

Mistral is known for permissive European open weights. It describes Mistral Large 3 as one of the best permissive open-weight models available and releases its models under Apache 2.0, alongside smaller sizes for lighter workloads.[4] The maker, Mistral AI, is a French company based in Paris.[5] The Apache license lets companies build on the weights with few strings attached.

Gemma, from Google, is the small, efficient counterpart to the larger Gemini family. Gemma models come in lightweight sizes meant to run on laptops, desktops, or your own cloud, including a compact 270M version, and they ship with open weights.[6] The family is the usual pick when you want something small enough to run close to where it is used.

Where to find these model families

References

  1. Llama 4 Scout 17B model card — Meta / Hugging Face
  2. Qwen3 — Qwen team, Alibaba Cloud
  3. DeepSeek-V3 model card — DeepSeek-AI / Hugging Face
  4. Introducing Mistral 3 — Mistral AI
  5. Mistral AI — Wikipedia
  6. Gemma 3 270M model card — Google / Hugging Face

Related