Mistral AI launched Mistral Large 4 on October 6, 2026 — a natively multimodal mixture-of-experts model in public preview via API, with open weights promised later this month.
2026-10-06 · by Moe Ameen
Mistral AI released Mistral Large 4 — nicknamed "Le Chonk" — on October 6, 2026, in public preview through its API. It is a natively multimodal mixture-of-experts (MoE) model: roughly 1.05 trillion total parameters with about 49 billion active per token, paired with a 1.6-billion-parameter vision encoder so it reads images as well as text. Mistral says it was trained on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe and supports more than 160 languages, including every official language of the European Union.
The model is available immediately via the API; Mistral has said it plans to release the open weights later in October 2026, though it has not confirmed the license they will ship under. Mistral lists a context window of up to one million tokens, but some third-party hosts are serving a smaller window in practice, so treat 1M as the vendor ceiling rather than a guaranteed served length. Pricing at launch was listed around $1.36 per million input tokens and $4.18 per million output tokens; Mistral's model card has since shown reduced rates closer to $0.68 input and $2.09 output per million tokens.
On its own numbers, Mistral pitches Large 4 as the strongest open-weight model developed in the US or Europe, citing 93% on Cybench, 61.7% on DeepSWE v1.1, and a 49.8% Coding Agent Index that it says edges out DeepSeek V4 Pro and Qwen3.8 Max, plus knowledge-work and visual-grounding scores it claims rival frontier closed models such as GPT-6-Astra. As with any vendor-reported launch benchmarks, independent evaluations will take time to confirm. One clarification for creators: Large 4 takes text and images in and writes text out — it generates no images, video, or audio.
The useful move the week a model like this ships is not to switch tools — it is to put its cheaper, longer-context drafting to work and then get the output in front of an audience. Mistral Large 4 will happily take a messy transcript, a research dump, or a week of notes and hand back a tight script or post. Then it stops. It renders no video, cuts no clip, designs no carousel, and schedules nothing.
That last mile is [Kompozy](/). Paste a Large 4 draft — or just an idea — into Kompozy and it generates the finished formats the model cannot: a [Persona Short](/glossary/persona-shorts) narrated by a face-locked HeyGen avatar, [Clipped Shorts](/) from a longer talk, a brand-exact [Carousel via HyperFrames](/glossary/hyperframes), Photo Posts, Quote Graphics, a Blog Article, and an Email Newsletter — 18 formats in all, each held to your voice by the [Persona Brief](/glossary/persona-brief). Then [Autopilot](/glossary/autopilot) reframes every clip to 9:16, 1:1, and 16:9 and schedules and publishes across eight social platforms plus blog and email behind a per-post review. Kompozy's own copy runs on managed Claude and OpenAI models, so you do not have to wire Large 4 in to use it — but if you are already drafting in Large 4, Kompozy is where that draft becomes a week of published, on-brand content.
Mistral Large 4, nicknamed "Le Chonk," is a natively multimodal mixture-of-experts language model from the European lab Mistral AI, released October 6, 2026. It has roughly 1.05 trillion total parameters with about 49 billion active per token, reads both text and images, writes text, and supports more than 160 languages.
It launched October 6, 2026 as a paid public preview through Mistral's API. Mistral has said it will release the open weights later in October 2026, but has not confirmed the license, so "open-weight" is the safer description than "open source" until the terms are published.
No. It is multimodal on input — it can take images and text — but its output is text only. It does not generate images, video, or audio. To turn a Large 4 script into finished video, carousels, or graphics you pair it with a content engine such as Kompozy.
At launch Mistral listed roughly $1.36 per million input tokens and $4.18 per million output tokens; its model card has since shown reduced rates closer to $0.68 input and $2.09 output per million tokens. Check Mistral's current pricing page before budgeting, as preview prices move.