Mistral Large 4 debuts: 1.05T parameters with 49B active, weights out October 27
赛博禅心 · wechat · 2026-10-06
Mistral released Mistral Large 4 in public preview: a fine-grained MoE with 1.05T total parameters and 49B active per token, a 1.6B vision encoder, native multimodality, and 1M context. Trained from scratch in two months on 4,000 Grace Blackwell chips in European datacenters.
Key preview scores: 62% on DeepSWE (beating GLM-5.3, DeepSeek-V4-Pro, Qwen3.8Max), 67% on Finch enterprise finance (tied with DeepSeek), 15% on Harvey legal agent tasks (best but still failing), 73% on DIOR-RSVG visual grounding, beating GPT-6 Astra.
Why open weights: cyber defense is the top use case — teams want to scan systems at scale without closed-model guardrails. Trained on 160+ languages, deployable on-prem or in EU regions; Mistral serves 125+ enterprises including Airbus, ASML, HSBC.
Pricing: $0.68/M input, $0.07/M cached, $2.09/M output. Weights drop October 27 under a custom license. The LeChonk name nods to the viral fake LeChatonFat meme from June.
Related event: Mistral Unveils Mistral Large 4, a 1T-Parameter Open Multimodal Model(42 posts)→
More from Models
- Google rolls out Nano Banana 2.1 with 4K output at $0.076 per image — testingcatalog · 2026-10-06
- Space Bunny Alpha lingers on OpenCode a few days longer than OpenRouter — sull · 2026-10-06
- Analysis: US Models Are Now Cheaper per Task Than Chinese Ones, Upending a Widely Held Assumption — sebkrier · 2026-10-06
- GPT-6.1 Sol spotted in Copilot 365 Premium, ChatGPT rollout likely next — koltregaskes · 2026-10-06
- Are models only improving at verifiable domains? AI stories winning prizes spark debate — erikphoel · 2026-10-06
- Reflection AI's upcoming open-weight model could bring the open-source crown back to the US — The AI Daily Brief · 2026-10-06