On Oct. 29, Mistral announced it had released the weights for an improved Mistral Large 4, also called Le Chonk. The company directs users to Hugging Face to download the weights and says they can customize the model with Mistral Forge and deploy it elsewhere. Unsloth's Hugging Face page describes Large 4 as a natively multimodal, 1 trillion-parameter mixture-of-experts model with 49 billion active parameters per token, or 52 billion counting embeddings and output layers.
Artificial Analysis reports that Large 4 at its high-reasoning setting scored 26 on the Artificial Analysis Intelligence Index v5, compared with 25 for Large 4 Preview. It used 82,000 output tokens per index task, versus 135,000 for Preview, a 39% reduction on that benchmark, the group reports. On its updated AA-Omniscience v1.1 benchmark, Artificial Analysis reports a 28% hallucination rate, down from Preview's 41%; Large 4's accuracy on that test was 26%.
Unsloth also lists its own quantized GGUF versions, including a 233 GB 1-bit variant. Its model page identifies the Mistral Ecosystem Open-Weight License as Large 4's license and says operating a Model-as-a-Service business requires a separate commercial license from Mistral. Artificial Analysis describes the same restriction.
Comments