Mistral Large 4 "Le Chonk": an EU-trained 1T open model with no published license

Mistral put Mistral Large 4 ("Le Chonk") into public preview on October 6, 2026. It is a trillion-parameter mixture-of-experts model trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers. On October 21, Mistral and NVIDIA announced a partnership on AI for industrial engineering, built on ML4. It includes new public benchmarks for agentic computer-aided engineering (CAE) and electronic design automation (EDA). Mistral's announcement page says "Weights drop end of this month".

I don't know of another open-weight model this large that was trained end to end in the EU on hardware the vendor runs itself. If you work in a regulated European industry, that is the reason to pay attention. The catch: as of today you cannot read the license, and you need a full 8-GPU Blackwell node to run it.

What actually shipped

Numbers that matter

Mistral's self-reported scores include DeepSWE v1.1 at 61.7%, Terminal-Bench 4.0 at 28.3% and Cybench at 93%. The 28.3% on Terminal-Bench is low for a 2026 frontier model, and Mistral doesn't hide it. On the independent side, explainx.ai reports an Artificial Analysis Intelligence Index of 38.4, which I could only confirm through secondary coverage.

Serving arithmetic, weights only, before KV cache and activations:

Mapped onto standard 8-GPU nodes, using vendor spec-sheet HBM sizes, the picture is this. On 8× H200 (141 GB each, 1,128 GB), FP8 weights take 1.05 TB and leave about 78 GB for KV cache across the node. That technically fits but isn't usable for 1M context or real batches, so plan on 16 GPUs. NVFP4 fits, but Hopper has no native FP4 math, so you pay to dequantize. On 8× B200 (180 GB each, 1,440 GB), FP8 leaves about 390 GB of headroom, enough for one node. NVFP4 leaves about 850 GB, which is the comfortable option.

The decode ceiling: at batch 1, each token reads about 52 GB at FP8. An 8× B200 node has roughly 64 TB/s of aggregate HBM bandwidth (8 TB/s per GPU), so the theoretical ceiling is around 1,200 tokens/s, minus expert-parallel communication overhead. At large batches, most experts are touched every step, and compute and all-to-all traffic become the limit.

Illustrative API cost (my workload): 2B input tokens a month, 80% of them cached, plus 200M output tokens.

One rented 8× B200 node costs more than that per month. Self-hosting is justified by residency and control, not price, unless you keep the node busy.

Until the file is published, "open weights" here means downloadable bytes and nothing more.

Technical meaning and trade-offs

The license is the whole story, and nobody has seen it. A "custom Mistral license" could be close to Apache 2.0, or it could look like the Kimi K3 license, which is MIT-style with a $20M revenue threshold for Model-as-a-Service providers and a branding clause above 100M MAU. It could also be a research-only Mistral license with a commercial carve-out. Each of those leads to a different decision for an EU bank or an OEM. Until the file is published, "open weights" here means downloadable bytes and nothing more.

Sovereignty applies to the weights, not automatically to the API. Training happened in the EU on Mistral-operated hardware, and there is an EU-operated deployment. Mistral's API also runs in "multiple regions worldwide". If residency is the requirement, pin the region in your contract. The default endpoint won't guarantee it. I've covered the broader picture in Mistral and EU sovereignty.

The CAE/EDA angle is a roadmap, not a result. Mistral and NVIDIA committed to creating benchmarks covering tool use, multistep correctness, run time and compute cost. The design is good, but there are no numbers yet.

My read: on the published numbers, ML4 is not the strongest open MoE. Kimi K3-class models score well above it on agentic coding. See trillion-scale open MoE compared. Its value is jurisdiction plus a 52B active path that fits on one Blackwell node at FP8.

What I'd do

Who should care: EU teams in defense, automotive, semiconductors, healthcare and the public sector, whose procurement rules out US- or China-hosted inference. Everyone else should treat ML4 as one more router option at $1.36/$4.18.

Read the license before downloading. Check four things: MaaS and revenue thresholds, field-of-use restrictions, attribution, and whether fine-tuned derivatives inherit the terms.

Plan for 8× B200 at FP8, or NVFP4 if quality holds. Budget KV cache explicitly if you want long context. Wait for an official NVFP4 checkpoint before quantizing it yourself; Mistral shipped one for Large 3.

Benchmark your own agent loop on the preview API now, before you commit hardware. Ignore the CAE/EDA story until it has scores.

What I couldn't verify

The active parameter count is 52B in Mistral's docs vs 49B in the NVIDIA post and VentureBeat, and I don't know which one is current. The license terms are completely unpublished; the words "custom license" come from VentureBeat, not from Mistral. On the weights date, VentureBeat says October 27, while Mistral only says "end of this month", and no ML4 repo exists on Mistral's Hugging Face org yet.

The Artificial Analysis score (38.4) and the promotional preview pricing come from an aggregator (explainx.ai), not from AA or Mistral directly. The benchmarks are almost all self-reported, and I found no CAE/EDA scores at all. The hardware HBM and bandwidth figures are vendor spec-sheet numbers, not measurements.

I'll revisit when the weights and LICENSE file land.

Sources

Primary: Mistral: Mistral Large 4 announcement; Mistral docs: Mistral Large 4; Mistral: Mistral and NVIDIA, AI for industrial engineering. Coverage: VentureBeat: Mistral debuts Large 4 Le Chonk; explainx.ai: Mistral Large 4 Le Chonk preview. Comparison: Hugging Face: Kimi K3 LICENSE (for comparison).