Mixtral 8x22B
Mistral AI · France · 2024
Mistral's largest Apache 2.0 model: 141B parameters stored, 39B used per token, 64K context and native function calling.
Mixtral 8x22B, published on 17 April 2024, scaled the sparse recipe of its predecessor to the largest model Mistral has ever released under a permissive licence. It stores 141 billion parameters and activates 39 billion per token, which keeps it faster than any dense 70B model while placing it above every other open-weight model of its moment on reasoning and knowledge benchmarks. Two additions separate it from Mixtral 8x7B beyond raw size. The context window doubled to 64,000 tokens, enough to hold a long contract or a codebase module in one prompt. More consequentially, function calling was native rather than bolted on: combined with the constrained output mode on Mistral's platform, it let developers wire the model into existing software as a component that returns structured results, not prose. It remained fluent in English, French, Italian, German and Spanish, with strong mathematics and coding. The licence was Apache 2.0 again, and Mistral released the base model alongside the instruction-tuned one specifically so that others could fine-tune it. This was the high-water mark of that strategy: from Mistral Large onward the company moved its biggest models to a research licence with paid commercial terms, and later flagships were never opened this way. Mixtral 8x22B is no longer part of Mistral's API line-up and no shutdown date was announced; independent providers still serve it and the weights stay downloadable.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!