Mixtral

Mixtral 8x7B was released by Mistral AI in December 2023 and quickly became one of the most significant open-source model releases, demonstrating that mixture-of-experts architecture could produce frontier-competitive performance at a fraction of the inference cost.

**The mixture-of-experts architecture:** Mixtral has 8 expert networks of 7B parameters each, but only 2 are activated for any given token. This means the model has 47B total parameters but runs with the compute of a 13B model — high quality at lower cost.

**Performance:** Mixtral 8x7B outperforms Llama 2 70B on most benchmarks and matches GPT-3.5 on many tasks, despite being faster and cheaper to run. Mistral's subsequent Mixtral 8x22B pushed performance further, competitive with GPT-4 on coding and reasoning.

**How to access:** - **API:** Available via Mistral's La Plateforme API (pay-per-token), with a free tier for testing - **Le Chat:** Mistral's consumer interface at chat.mistral.ai uses Mixtral under the hood - **Self-hosted:** Models available on Hugging Face; run locally via Ollama, LM Studio, or vLLM - **Third-party:** Available through Perplexity, Poe, and many AI API aggregators

**Strengths:** Strong coding performance. Multilingual by design (French, German, Spanish, Italian in addition to English). Fast inference. Open weights with a permissive Apache 2.0 license for the 8x7B model.

**Mixtral 8x22B:** The larger variant (released April 2024) is significantly more capable, with 141B total parameters (39B active), competitive on complex reasoning tasks and coding with models far above its weight class.

**Alternatives:** Meta's Llama 3 70B is the primary open-source competitor. Google's Gemma 2 27B is smaller but competitive. For API access with similar pricing, Groq offers extremely fast Mixtral inference.