Mistral Large 4: Europe's Open-Weight Model for Enterprise AI
October 6, 2026: Mistral AI today launched a public API preview of Mistral Large 4 (ML4), a 1-trillion-parameter natively multimodal model the Paris-based company positions as the most capable open-weight AI system built outside China. The model’s in-house nickname is “le Chonk.”
The release is the first milestone on Mistral’s €3 billion Series D, the largest equity round ever raised by a European technology company. It lands as Western AI labs race to close a capability gap that Chinese open-weight models from Qwen, DeepSeek, Kimi, and GLM have defined for the past year.
Weights drop by October 27. Until then, the API is live in Mistral Studio.
What Mistral Large 4 Is
ML4 is a granular Mixture-of-Experts (MoE) model with 1.05 trillion total parameters and 49 billion active parameters per token. That active footprint cuts inference compute significantly, though enterprises should note that all 1.05 trillion parameters must still be held in memory for deployment.
Key technical specifications:
| Specification | Value |
|---|---|
| Total parameters | 1.05 trillion |
| Active parameters per token | 49 billion |
| Vision encoder | 1.6 billion parameters |
| Context window | 1 million tokens |
| Training languages | 160+ (all official EU languages included) |
| Training infrastructure | 3,800 NVIDIA Grace Blackwell GPUs, Mistral’s European data centers |
| API pricing (Standard) | $0.68 / M input, $2.09 / M output |
| Weights release | End of October 2026 |
The model is natively multimodal: it accepts text and images in the same context. A 1.6-billion-parameter vision encoder handles visual input, including satellite imagery, technical drawings, documents, and charts.
Benchmark Performance
Mistral positions ML4 as the leading open-weight model from any Western lab on aggregate benchmarks. The Artificial Analysis Intelligence Index places it at 38 points, a major jump from Mistral Large 3 at 9 points. For reference, Claude Opus 5.5 (Max) leads that index at 58. The gap to frontier closed models is real. On most individual benchmarks, leading Chinese open-weight models from DeepSeek and Kimi still score higher.
The table below compares ML4 against key open-weight peers on task-specific benchmarks. All figures come from Mistral’s published announcement and independent evaluators cited there.
| Benchmark | Mistral Large 4 | Nearest Open-Weight Peer | Notes |
|---|---|---|---|
| DeepSWE v1.1 (coding agent) | 61.7% | DeepSeek V4.1 Flash: 74.2% | Kimi K3 at 68.0% also higher |
| AutomationBench (657 workflows) | 59.9% | GLM-5.3: 62.2% | Gmail, Sheets, Slack, Salesforce |
| CyberGym-E2E (patch vulnerability) | 82% | MiMo-V2.6-Pro: 79% | Highest of any model tested |
| Cybench (40 CTF challenges) | 93% | Among top open-weight scores | |
| Dense200 visual grounding | 42% | GPT-6 Astra (closed): 41% | Open-weight lead |
| Harvey Legal Agent Benchmark | 15.8% | Kimi K3: 12.9% | Best open-weight result |
| Lakera B3 (attack resistance) | 93.3% | GLM-5.3: 93.3% (tied) | Prompt injection |
| FinWorkBench (finance tasks) | 67.4% | DeepSeek V4 Pro: 67.4% (tied) | |
| Artificial Analysis Intelligence Index | 38 pts | Claude Opus 5.5: 58 pts | Aggregate across domains |
One result deserves particular attention: ML4 scores 82% on the reproduce-and-patch benchmark, the highest of any model. Claude Opus 5.5 and GPT-6 Astra score near zero on the same test, not because they lack the capability, but because they refuse to perform it. This is not a performance gap; it is a policy gap, and Mistral has built its enterprise pitch around it.
The Cybersecurity Angle, Explained
Defending a system often starts with proving that a vulnerability is real before writing a patch. That requires running code, tracing memory, and demonstrating exploitation. Most closed model providers refuse these tasks as potential offensive use.
Mistral’s argument is straightforward: if a defender cannot use an AI model for legitimate security research, the only beneficiaries of that refusal are attackers who jailbreak the same closed models anyway.
ML4 handles this class of work natively. On Cybench, a set of 40 challenges drawn from real security competitions, it solves 93% of tasks. On the Artificial Analysis Cyber Index, it ranks among the top five models globally and leads open-weight models from outside China by a wide margin.
Critically, open weights make self-hosted deployment possible. A financial institution or government agency can run ML4 on its own servers, apply its own moderation policies, and analyze sensitive code without any of that data leaving the organization’s perimeter. Until weights drop, Mistral is red-teaming ML4 with cybersecurity firms, vetted partners, and government agencies who receive a version with expanded cyber capabilities and reduced moderation.
For enterprises evaluating AI for security operations, this is the most significant open-weight release of the year.
Enterprise Verticals: Finance, Legal, and Manufacturing
Mistral trained ML4 in collaboration with enterprise customers across finance, manufacturing, logistics, pharmaceuticals, shipping, and the public sector. Those customers used the same training and reinforcement learning environment Mistral sells through its Forge platform.
Finance. ML4 scores 67.4% on FinWorkBench, tied with DeepSeek V4 Pro and ahead of GLM-5.3. An independent evaluation by Vals.ai places it above GPT-6 Astra on financial tasks. Mistral cites 125+ enterprise customers including HSBC and ASML.
Legal. On Harvey’s Legal Agent Benchmark, ML4 scores 15.8%, ahead of every other open-weight model. The absolute numbers are low across all models on this benchmark, which means neither ML4 nor any competitor is ready for unsupervised legal work. The result is still meaningful: it establishes ML4 as the best open-weight option for legal workflow assistance today.
Manufacturing and science. ML4 can convert technical drawings into CAD models, analyze semiconductor designs, and handle multi-step scientific simulations including chemistry modeling. In an internal expert evaluation against GLM-5.3, Mistral’s annotators preferred ML4 in STEM (68% win rate) and CAD (62% win rate).
Vision. Earth observation is an emerging enterprise application. ML4 scores 73% on DIOR-RSVG, a remote-sensing benchmark for locating objects in satellite imagery, compared to 68% for GPT-6 Astra. Mistral specifically calls out insurance (storm damage assessment), utilities (power line inspection), and agriculture (crop health monitoring) as target use cases.
European Sovereignty as a Product Feature
ML4 was trained entirely in Mistral’s own European data centers, on 3,800 NVIDIA Grace Blackwell GPUs running about 10 megawatts of power over approximately two months. The preview API runs on that same infrastructure.
Once weights are public, organizations can self-host under their own governance. Mistral also operates a European API region that processes data exclusively under EU law, independently of other cloud providers.
This matters in regulated industries. Financial services firms, health-care providers, and government agencies face data residency requirements that make US-hosted models difficult or impossible to use for certain workloads. The enterprise AI sovereignty conversation has been accelerating through 2026; ML4 gives European organizations a credible on-premise option that was not available before.
Guillaume Lample, Mistral’s co-founder and chief scientist, framed the political dimension directly: “The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks.”
What Comes Next
October 27, 2026. Weights release under an open license. Mistral Large 3 shipped under Apache 2.0; the same is expected for ML4, though terms have not been confirmed. Quantized FP8 and NVFP4 builds will also be released, reducing hardware requirements for inference.
Specialized models. ML4 is not a finished product; it is a base. Mistral plans a new generation of fine-tuned vertical models built on ML4, covering the industries where it already has customer relationships.
Ongoing training. Mistral noted that the reinforcement learning run behind the preview is still in progress and “showing no signs of saturation.” Benchmark numbers will continue to improve before the weights ship.
Infrastructure scale-up. Mistral is using its €3B Series D to expand European data center capacity, targeting 200 megawatts of compute by end of 2027. ML4 is the first model to come out of that investment.
What Enterprise Leaders Should Do Now
Three concrete actions make sense before weights drop:
-
Run the API preview on your highest-value use case. The preview already reflects significant capability, and evaluation against a specific workflow reveals more than aggregate benchmarks. Focus on cybersecurity, legal document review, financial analysis, or multilingual workloads where ML4’s differentiation is clearest.
-
Plan hardware ahead of the weights release. A common mistake with MoE models is sizing infrastructure for the active parameter count (49B) rather than the full parameter count (1.05T). All weights load into memory at startup. This is not a model for a single workstation; it requires serious GPU infrastructure or a coordinated multi-node deployment.
-
Evaluate enterprise AI cost structure. At $0.68 per million input tokens, ML4’s preview pricing is competitive with DeepSeek V4 Pro. Organizations already running cost models for LLM workloads should slot ML4 into those models before weights land and API pricing potentially shifts.
ML4 does not displace frontier closed models at the aggregate intelligence level: the Artificial Analysis index gap between 38 and 58 is real. But for organizations that need European data residency, on-premise deployment, or AI that will actually perform security research without refusing, it is the most significant open-weight release from a Western lab in 2026.
Sources: Mistral AI official announcement (October 6, 2026); The Next Web (October 6, 2026); The Decoder (October 6, 2026); CNBC (October 6, 2026).