On this page
Last Updated: October 7, 2026
Key Takeaways
- Mistral Large 4 is a 1.05T parameter open-weight MoE model with 52B active parameters, a 1M token context window and a 1.6B vision encoder, live in public preview as of October 6, 2026 (Mistral AI).
- It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, making it the most significant European AI sovereignty play of 2026 (Mistral AI).
- Cybersecurity is its strongest vertical: 82% on Artificial Analysis' vulnerability reproduction test, the highest of any model, and 93% on Cybench (Mistral AI, Artificial Analysis Cyber Index).
- On agentic coding it posts a 49.8% Coding Agent Index, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max (Mistral AI).
- It beats GPT-6-Astra on Dense 200 visual grounding (42% vs 41%), one of the few open models to surpass a frontier closed model on any benchmark (Mistral AI).
- Pricing is $0.14 per million input tokens and $0.68 per million output tokens, roughly a quarter of comparable frontier API pricing (Mistral docs).
What is Mistral Large 4?
Mistral Large 4 (ML4, internally "le Chonk") is Mistral AI's new flagship frontier model: a natively multimodal, trillion-parameter model with 52 billion active parameters, a granular Mixture-of-Experts architecture, and a 1M token context window. It went live in public preview on October 6, 2026, with open weights promised by the end of the month.
Three numbers matter here. 1.05 trillion total parameters gives it the capacity of a frontier lab's biggest model. 52 billion active parameters per token means you get frontier performance at roughly a 20th of the inference cost of a dense model that size. And the 1M context window means it can hold your entire codebase, a full legal discovery dump, or a season of customer transcripts in memory. All in a single model, all trainable on European soil.
How was Mistral Large 4 trained?
ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. That matters. This is the first serious frontier-class model trained entirely in Europe, on European infrastructure, under European law. A significant share of the training data covered more than 160 languages, including every official EU language.
Mistral's reinforcement learning recipe is the real unlock. At their 3K GPU scale, a single training run produces roughly 33 billion tokens per day, of which about 16 billion are trainable completion tokens after filtering and masking. RL generation and training environments are composable, mixing chat, safety alignment, factuality, tool use, code sandboxes, web search and external APIs in one run.
After the European sovereignty story, this is the second most important thing about ML4. The €3 billion Series D (largest equity round ever raised by a European technology company) means more compute is coming online monthly, and Mistral is explicitly signaling the model will keep improving.
Benchmarks that actually matter
Agentic coding is the metric most enterprise buyers care about, and ML4 delivers a 49.8% Coding Agent Index (61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4), placing it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max among open-weight models.
| Benchmark | ML4 | vs Field |
|---|---|---|
| Artificial Analysis Cyber Index | 82% vuln-repro | Highest of any model globally |
| Cybench (40 security challenges) | 93% | Among the top open-weight scores reported |
| Coding Agent Index (combined) | 49.8% | Ahead of DeepSeek V4 Pro 0813, Qwen3.8 Max |
| Dense 200 (visual grounding) | 42% | Above GPT-6-Astra (41%) |
| Lakera B3 AI Security Benchmark | 93.3% attack resistance | No higher score found among competitors |
| KORA safety benchmark | 1.691 / 2.0 | Highest among open-source models |
Safety deserves its own paragraph. ML4 saturated Mistral's benchmarks on robustness to indirect prompt injections, posting 93.3% resistance on Lakera's public B3 AI Security Benchmark, the highest score Mistral found among competitors. On the KORA benchmark it scores 1.691 out of 2 ("Exemplary"), again the highest among open models. Tighter refusal behavior than any previous Mistral model, measured via JailbreakBench, StrongREJECT and AgentHarm, plus that safety-per-cyber-capability pairing is what makes it usable in regulated environments.
Mistral Large 4 pricing
Per Mistral's official docs, ML4 costs $0.14 per million input tokens and $0.68 per million output tokens, with cached input at $0.07. Batch mode halves everything again: $0.07 input, $2.09 output per million tokens. That is the frontier-capability tier of pricing at a mid-tier model cost. Compare it to the $4 to $15 per million output tokens typical of closed frontier APIs in 2026 and the ROI story writes itself.
In our dual DGX Spark testing and client deployments across Australian professional services firms, we routinely see 40% to 70% total cost reduction when a workload moves from a closed frontier API to a well-chosen open-weight model hosted on owned infrastructure. ML4's combination of 52B active parameters, EU hosting options and sub-$1 output pricing pushes that math further than anything we have tested this year.
Who Mistral Large 4 is for
ML4's sweet spot is organizations that need frontier capability without frontier vendor lock-in. Concretely: cybersecurity teams running vulnerability research and incident response on their own terms, with model weights they control. Financial services and law firms that need a model that reads contracts, edits spreadsheets and reasons over filings without shipping customer data offshore. European enterprises with data residency requirements (and Australian businesses that prefer their AI not route through US hyperscalers). And the growing cohort of builders running agent fleets where per-token economics determine viability.
It is less ideal for teams who need the absolute fastest time-to-market with zero ops overhead, or for tiny ventures without the engineering bandwidth to fuzz-test deployments. For that latter cohort, the API preview at $0.14/$0.68 makes ML4 worth piloting even fully managed.
How Flowtivity reads it
Mistral Large 4 is the first model that makes the "sovereign AI" pitch economically rational rather than merely patriotic. You get frontier-class capability, frontier-class safety metrics, US-grade open weights on your own terms, and European-grade pricing. In 2026 the real question for enterprises is not "should we use a frontier model" but "which frontier model can I run under my own roof, under my own policies, at a price that survives procurement review."
For Australian businesses, the answer increasingly looks like: the one trained in Europe, open-weighted, and priced at a quarter of the closed alternatives. That question now has a strong answer.
FAQ
Q: What is Mistral Large 4?
A: A 1.05T parameter open-weight multimodal MoE model with 52B active parameters and a 1M context window, in public preview as of October 6, 2026, with open weights by end of October.
Q: How much does Mistral Large 4 cost?
A: $0.14 per million input tokens, $0.68 per million output tokens, cached input $0.07 (Mistral docs).
Q: Is Mistral Large 4 open weight?
A: Yes, weights release by end of October 2026 after a red-teaming phase.
Q: When did Mistral Large 4 launch?
A: Public preview launched October 6, 2026, trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters.
One email a month, no noise
Practical AI notes for Australian businesses. Unsubscribe anytime.
