Mistral is releasing Large 4's open weights: a 1.05T MoE, 52B active
Today's #1 on Hacker News: Europe's frontier open model is already callable at ~$0.70/$2.10 per 1M tokens, with weights landing on Hugging Face this week.
Copy markdown
A 1.05T MoE, 52B active, natively multimodal
Mistral Large 4 is a mixture-of-experts flagship: 1.05T total parameters but only 52B active per token, plus a 1.6B vision encoder for native image input. On Mistral's own benchmarks it posts a ~49.8% Coding Agent Index and solves 93% of Cybench challenges — a frontier-class model you can put under your own control instead of renting.
Callable today at ~$0.70 in, ~$2.10 out
The hosted model is live now on Mistral Studio (La Plateforme) at about $0.68 per 1M input and $2.09 per 1M output tokens — a promo rate roughly half the ~$1.36/$4.18 list. Prototype against the API today, then swap to your own weights when the download lands: it's the same model either way.
Weights hit Hugging Face this week — check the license and the hardware
Mistral says the open weights drop by end of month on Hugging Face. Early reports describe a modified-MIT-style license that lets you run, fine-tune and ship the model, with one carve-out: reselling it as a managed service needs a separate commercial license. Budget the GPUs too — the 1T weights run ~1TB at FP8 and ~0.5TB at 4-bit, so this is a multi-GPU deploy, not a laptop.