Mistral AI’s October 6, 2026 announcement introduces Mistral Large 4, which the post jokingly calls “le Chonk,” as a public preview you can call today through the preview API on Mistral Studio. The bigger news for Alberta is a promise in the same post: open weights. For energy, industrial, and public-sector operators who want frontier-class capability on hardware they control, that changes the planning conversation, even though the weights, a hardware spec, and full architecture details are not out yet.
What Mistral says it launched
The primary is Mistral’s own post. It calls ML4 “a 1 trillion-parameter natively multimodal model with 49 billion active parameters” and Mistral’s largest and most capable model to date. The preview API is live on Mistral Studio. Mistral says ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe, and that the public preview is served on that same infrastructure. It adds that a significant share of the training data was multilingual, spanning more than 160 languages. On the weights, the post is consistent: “We will release the weights by the end of the month.” Some press coverage reports a specific release date and a licence. Mistral’s pages that we read state neither, so we do not print them as Mistral’s.
“ML4 pairs top-tier cyber performance with open weights and self-deployment, giving organizations both the capability and the autonomy to run advanced security work under their own policies.”Mistral — Introducing Mistral Large 4, October 6, 2026
Until the weights ship, Mistral says it is red-teaming the model in real-world settings with cybersecurity leaders, vetted partners, and state authorities, who will access the same model with reduced moderation and expanded cyber capabilities. That is a restricted access group, not the public preview. Mistral also says the model will be available across multiple regions worldwide, including a European deployment it operates end-to-end under European law. The post does not name a Canadian region.
Size, context, and preview pricing from Mistral’s docs
Mistral’s model page tags the release Public Preview, Open, and v26.10. It describes a granular Mixture-of-Experts architecture with 49B active parameters, 1.05T total parameters, and a 1.6B vision encoder, plus a 1M context window. The blog rounds the size to one trillion; the docs give the more precise 1.05T. Features listed include structured outputs, function calling, document Q&A, batching, and agents and conversations endpoints. Prices are shown per million tokens, with a list price struck through and a lower preview price beside it. Both are below. The docs do not say how long the preview price lasts, and we have not converted anything to Canadian dollars.
| Token | Preview price shown | List price (struck through) |
|---|---|---|
| Input | $0.68 | $1.36 |
| Cached input | $0.07 | $0.14 |
| Output | $2.09 | $4.18 |
Mistral’s benchmark claims, and what is still missing
Every benchmark figure here is Mistral’s claim, not an independent Alberta test. Mistral says that on critical enterprise workloads, including cybersecurity, finance, and law, it finds ML4 to be “state-of-the-art among open models.” In its cybersecurity section, Mistral reports that ML4 solves 93% of the challenges in Cybench. In its safety section, Mistral reports that ML4 resists 93.3% of attacks on Lakera’s public B3 AI Security Benchmark. Mistral says more benchmarks, architecture details, and its post-training methodology will come with the weights.
What is missing matters more for self-hosting. Mistral says that for organisations that need sovereign, auditable AI for security operations, the model “will be able to run on private cloud or on-premise.” Neither the post nor the docs publish a self-host hardware spec, a GPU count, or memory requirements, and we do not guess one. Sizing a 1.05T-parameter deployment is a serious infrastructure decision, and it should wait for Mistral’s published requirements and licence terms.
What Alberta operators should plan now (our advice)
What follows is our advice, not Mistral’s. Energy, industrial, utility, and public-sector teams in Alberta that cannot send sensitive data to a foreign API now have a frontier-class open-weight option to plan around. Plan it in three steps. First, pilot on the preview API with non-sensitive or synthetic data only; Mistral says the preview is served from its European infrastructure, not from Canada. Second, write down the one workflow you would self-host, such as engineering-drawing review, document Q&A over public standards, or security-operations triage, and the evaluation set you will judge it on. Third, size hardware only after Mistral publishes self-host requirements and licence terms.
Then compare the result with a managed private platform. Our Cohere North 2 briefing covers an air-gapped, on-prem path with admin spend caps. Define the data boundary first with private AI security work; for plant and field use cases, see industrial AI Alberta. An AI consulting Alberta engagement can turn the pilot results into a build-or-buy decision.