Mistral returns with Large 4: a 1T-param 49B-active multimodal MoE, 1M context
Xianbao_QIAN · x · 2026-10-29
Mistral ships Mistral Large 4 (Mistral-Large-4-1T-A52B): a multimodal MoE with 1T total / 49B active params, sparse MLA attention, a built-in DSpark drafter, tool calling and reasoning.
- 1,048,576 token context window
- FP8, NVFP4 and AMD MXFP4 checkpoints on Hugging Face
- On 8x GB300 with DSpark speculative decoding, a TP4 setup reaches 241 output tok/s/GPU at 32k input
Related event: Mistral Open-Sources Trillion-Parameter Flagship Mistral Large 4(27 posts)→
More from Models
- Ex-OpenAI safety VP Miles Brundage: Claude Duet as a separate app is not ideal — Miles_Brundage · 2026-10-30
- Phonic 1 Hits 0.61s Time-to-First-Audio, Fastest of the Top 5 Speech Models — ArtificialAnlys · 2026-10-30
- Phonic 1 Costs $4.20/Hour of Input Audio, Undercutting Grok and Most GPT-Live Tiers — ArtificialAnlys · 2026-10-30
- Phonic 1 Scores 98.9% on Big Bench Audio, Tops Tau Voice Agentic Benchmark — ArtificialAnlys · 2026-10-30
- Phonic 1 Leads Tau Voice Agentic Benchmark at 79.9%, 5.4 Points Clear — ArtificialAnlys · 2026-10-30
- Phonic 1 Tops Artificial Analysis Speech-to-Speech Index at 87.1, Edging Grok Voice 3.0 — ArtificialAnlys · 2026-10-30