The Mistral model family is the catalog of French lab Mistral AI, split between open-weight releases that anchored the efficient-model movement and closed commercial flagships. The family's early releases made "small but overtrained" a deliberate strategy rather than a budget compromise.1

Family tree

Generations

Mistral 7B (September 2023) beat 13B-class rivals with sliding-window attention and careful data work, becoming the community's default base model under Apache 2.0; Hugging Face's Zephyr demonstrated its post-training headroom with a direct-preference recipe.3 The company's release rhythm became its signature: weights dropped as magnet links on X before any announcement blog, a gesture the open community read as cultural alignment. Mixtral 8x7B (December 2023) was the first widely deployed open sparse mixture-of-experts model, matching or beating GPT-3.5-era systems at a fraction of inference cost, as reported and independently reproduced. The commercial Large line and specialist models carried the company into API competition with OpenAI and Anthropic, while periodic open releases maintained its standing against Meta AI's Llama and the DeepSeek line on suites like MMLU. The split licensing model, Apache 2.0 for community releases and research or commercial terms for flagships, became the template other European and American labs cited when hedging between openness and revenue.12

See also

References


  1. Jiang, A., et al., "Mistral 7B," arXiv:2310.06825, October 2023. ↩↩

  2. Jiang, A., et al., "Mixtral of Experts," arXiv:2401.04088, January 2024. ↩

  3. Tunstall, L., et al., "Zephyr: Direct Distillation of LM Alignment," arXiv:2310.16944, October 2023. ↩