Mistral Large 4 "Le Chonk" Review: 1T Parameters, No Weights Yet

Mistral Large 4 "Le Chonk": a trillion-parameter preview with a licence nobody has read

On October 6, 2026, Mistral AI opened a public preview of Mistral Large 4. Its internal nickname is "Le Chonk," and the company used that name in the headline of its own launch post. The pitch is a frontier model you will eventually be able to download. The catch is the tense: as of October 12, you cannot download it, and no licence exists.

Here is what actually shipped, what it costs, and why the release pattern matters more than the model.

The specification

~1.05 trillion total parameters, ~49 billion active per token — roughly 4.9% activation
Natively multimodal: text and image in, text out, with a 1.6B-parameter vision encoder
1-million-token context window
160+ languages, including every EU official language
Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacentres — not fine-tuned on a Chinese open-source base, which Mistral emphasises

That last point is the commercial argument. Mistral is positioning itself as the European alternative for buyers who cannot put workloads on US-hosted closed APIs or Chinese-derived weights.

Pricing: undercutting the frontier on output

Mistral's published list price is $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14.

For comparison, both Gemini 4 Argon (introductory) and GPT-6.1 Sol sit at $2 / $10. Mistral undercuts Gemini on input by about 32% and on output by well over half. Among non-Chinese releases, that is the cheapest frontier-tier rate card published this month.

Two caveats matter more than the headline price. First, Mistral's own docs card currently displays $0.68 / $2.09 — exactly half — with no explanation of the discount or its duration. Second, this is preview pricing. Nothing has been committed for general availability.

Benchmarks: strong security case, weak general case

This is where you have to read two sets of numbers and notice that they disagree in direction.

Mistral's own published figures:

Benchmark

Score

Cybench	93%
CyberGym-E2E	82%
DeepSWE v1.1	61.7%
Terminal-Bench 4.0	28.3%

Independent measurement tells a different story about general capability. Artificial Analysis scores it 38 on its Intelligence Index — level with GPT-6 Luna at maximum reasoning effort, just behind DeepSeek V4.1 Flash at maximum effort (39), and 25th out of 25 tracked models. Its Cyber Index score of 50 matches GLM-5.3-Flash and trails MiMo-V2.6-Pro's 56.

Read those together and the picture is consistent: this is a model with a genuine specialism in vulnerability discovery and patching, and a respectable multilingual document capability, sitting well outside the general-capability top tier.

The Terminal-Bench number deserves separate attention. 28.3% against Claude Sonnet 5.5's 63.6% and GPT-6.1 Sol's 56.1% in the same week means agentic coding is not this model's lane. If you were hoping a trillion-parameter European model would double as a coding agent brain, the measurement says otherwise.

The open weights are promised, not shipped

Mistral's announcement says: "Weights drop end of this month." Reuters reported a date of October 27. Mistral itself only commits to the end of October.

As of this writing, no licence has been published. That is the decisive detail, and it is why a large share of the coverage you will find describes something that does not exist yet.

Three things are still open:

Licence terms. Mistral Large 3 shipped under Apache 2.0. VentureBeat reports Large 4 will use a custom Mistral licence instead. If you are planning commercial self-hosting, that difference is the whole decision.

The final checkpoint. The reinforcement learning run behind the preview is still in progress. The weights that eventually ship will not be identical to the preview API.

Red-teaming. Mistral says the weights will follow red-teaming with security partners and state authorities.

If your roadmap assumes self-hosting Mistral Large 4, you are currently assuming three unannounced facts.

The self-hosting arithmetic

Even once the weights land, running this thing is a serious infrastructure commitment:

~525 GB at INT4
~1.05 TB at FP8
~2.1 TB at BF16

For single-node production, the credible configuration is an 8× B200 server. This is not a consumer-GPU or workstation model, and the 4.9% activation rate does not change that — a mixture-of-experts architecture reduces compute, not memory. The weights still all have to be resident.

From an operator's standpoint, the 1M-token context is the trap. Filling that window through a self-hosted node costs real money in GPU hours, and for most workloads an API call will be cheaper than amortising eight B200s.

Where it is actually worth testing

This is not a general-purpose upgrade. It is a candidate for specific workloads:

Vulnerability discovery and remediation — Cybench 93% and CyberGym-E2E 82% are the vendor's numbers, but the specialism holds up in independent cyber evaluations

Multilingual document processing — 160+ languages with EU coverage is unusual and directly useful for European regulatory work

Legal drafting — Mistral reports one strong legal-drafting result

Mixed image-and-text document pipelines — native multimodality means scanned pages and text run through the same path

Where it is not worth testing: agentic coding, or high-volume long-context reasoning. Both already have stronger options at this price point.

The real story: announcements stopped equalling availability

The Mistral launch is one data point in a pattern that defined October 2026.

Google announced Gemini 4 Argon on September 30, scoring 52.6 on the Intelligence Index — level with GPT-6 Astra — and made it available only to vetted cyber defenders through its Fairwind Program, with no general availability date
OpenAI's GPT-6.1 Astra was reported cancelled outright after internal safety tests failed
Mistral Large 4 has a live API and promised weights
The eight models released October 1–3 were all specialists, none of them a frontier generalist

Three of the four cases are models you cannot simply buy and deploy. The capacity exists; the permission does not.

The engineering takeaway is concrete. Keep a second model qualified for every critical path. Treat vendor "coming weeks" language as an estimate, not a commitment. Prefer architectures where the router can be swapped without a redeploy. And when a launch headline says "open weights," open the model card and look for a licence file before you write it into a plan.

Bottom line

Mistral Large 4 is a 1.05T-parameter, 49B-active multimodal MoE covering 160+ languages with a 1M-token context
At $1.36 / $4.18 per million tokens it is meaningfully cheaper than Gemini 4 Argon and GPT-6.1 Sol on output
Its measured general intelligence sits in the bottom quartile of tracked models; its strength is cybersecurity and multilingual document work
The weights are not out. No licence, no final checkpoint, no published terms — the preview API is all you can use today
Self-hosting needs ~525 GB at INT4 and an 8× B200-class node in practice
The strategic lesson of October 2026 is that a named next-generation model is no longer a routed one