Mistral Large 4 Is Here: A 1.05-Trillion-Parameter Open-Weight Model You Can Actually Host Yourself

The most interesting AI model announced this week is the one you are allowed to run on your own servers.

On 6 October 2026, Mistral AI opened a public preview of Mistral Large 4 — unofficially ML4, officially le Chonk — a 1.05-trillion-parameter mixture-of-experts model of which only 49 billion parameters activate per token, with a 1.6B vision encoder making it natively multimodal. Unlike every other frontier-scale model of the moment, Mistral says the weights will be released, free, by the end of this month.

What actually happened

The scale numbers are the headline. ML4 is Mistral's largest model to date: 1.05T total parameters, 49B active, granular MoE, multimodal input, per the official model card. The practical consequence of a 49B-active footprint is serving cost — you pay for 49B parameters of work, not a trillion — while the trillion-parameter base carries the knowledge. Mistral is serving the preview from its own infrastructure at $1.36 per million input tokens and $4.18 per million output tokens.

It was trained in Europe, on Mistral's own metal. The company says ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacentres, and that the public preview is served from that same fleet. A significant share of the training data was multilingual — more than 160 languages, including every official language of the European Union. Availability is planned across multiple regions, including a European deployment Mistral operates end-to-end under European law. Mistral confirmed the launch on X with the same spec sheet: 1T parameters, natively multimodal, 49B active.

The cybersecurity results are the part worth reading twice. On the Artificial Analysis Cyber Index — an independent evaluation of how well models find and fix real security flaws — Mistral places ML4 among the top five models globally, and first of any model on a test that asks the model to reproduce a real vulnerability in open-source software and then patch it, at 82%. On Cybench, a set of 40 exercises drawn from security competitions, ML4 solves 93%. Mistral notes that several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on that same test because they decline to perform it.

Coding and agent scores are competitive, and vendor-reported. ML4 posts 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4.0, for a combined Coding Agent Index of 49.8% that Mistral says sits ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. In a blind human evaluation by Surge AI on coding quality it ranked second of five models at 3.74/5, behind Claude Opus 5 at 4.22 and ahead of GLM-5.3 and Kimi K3. On AutomationBench — 657 business workflows across Gmail, Google Sheets, Slack and Salesforce — it scores 59.9%, and on AA-Briefcase, which measures long-horizon knowledge work such as spreadsheets, slides and PDFs, it reaches 1,393 Elo. On visual grounding it edges GPT-6-Astra on Dense 200, 42% to 41%.

Under the hood, the RL run is still going. Mistral says the reinforcement-learning run behind this preview is trained on roughly 3,000 GPUs at about 33 billion tokens per day, of which around 16 billion are trainable completion tokens after filtering — and that the run has not saturated yet. The preview is therefore a moving target: expect the model to improve before the weights land. It is also being red-teamed in real-world settings alongside cybersecurity leaders, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capability. All of it is funded by Mistral's €3 billion Series D, which the company describes as the largest equity round ever raised by a European technology company.

Why this matters for Indian SMBs and mid-market teams

• Self-hosting changes the economics, not just the ideology. Weights due by the end of October means an Indian firm can run a frontier-class model inside its own VPC or on-premises, on an opex curve you control, instead of a per-token bill that repricing can move overnight. For teams handling payroll, health, or financial records, "the data never leaves our network" is also a much simpler answer to DPDP Act obligations than a vendor's data-processing addendum.
• Refusal-free defence is a real buying criterion. ML4's Cybench score is high precisely because it does the vulnerability-reproduction work that safety filters on closed models block. Any Indian services firm running vulnerability assessments, malware triage or incident response knows the frustration of a capable model declining the task mid-incident — a self-hosted model under your own policy removes that failure mode.
• The 49B-active architecture is what makes the price honest. At $1.36/$4.18 per million tokens in preview, ML4 sits in the mid-tier price band while claiming frontier-adjacent scores — and because only 49B parameters activate, self-hosting it is not a datacentre-scale fantasy the way a dense trillion-parameter model would be.
• Treat every benchmark here as vendor-reported. These numbers come from Mistral's own release page, the model is still in preview, and the RL training is unfinished. Validate against your own task set — your invoices, your tickets, your contracts — before you build a roadmap on them.

Where Guruji Corporation fits in

Open weights only matter if someone wires them up properly: quantisation, retrieval, evaluation against your real documents, and a monitoring layer that tells you whether quality drifted after a model swap. We build that stack for Indian SMBs and mid-market teams — AI-assisted document pipelines, support automation and internal assistants that can run on a hosted API today and move to your own infrastructure the week Mistral's weights drop, without a rewrite. Our evaluation harness runs your own tasks against both, so the migration decision comes back as a number, not an opinion.

If you are already paying per token for AI in October, the useful conversation this week is what your workload would cost on a model you host yourself. Talk to us and we will benchmark your real tasks against ML4's preview API, price both options at your volume, and tell you plainly whether October's weight release is a switch worth making or a distraction.

Sources

• Introducing Mistral Large 4 — Mistral
• Mistral Large 4 model card — Mistral documentation
• Mistral AI on X: Meet Mistral Large 4, aka Le Chonk