Mistral launched the public preview of Mistral Large 4 through the Mistral Studio API on October 6, 2026. The company describes it as a natively multimodal model that brings instruction following, reasoning and agentic capabilities together in one system. Mistral plans to release its weights at the end of October.
Coding: agent benchmarks and blind code evaluation
For coding agents, Mistral reports scores of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4.0. Its combined Coding Agent Index score is 49.8%, which the company says puts Large 4 ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. Artificial Analysis evaluated these three results privately before the test harness was publicly released.
A blind evaluation of code quality conducted by Mistral with Surge AI offered another perspective. Professional evaluators scored outputs on a scale of 1–5 without knowing which model produced them. ML4 Preview finished second with 3.74, behind Claude Opus 5 at 4.22; GLM-5.3 scored 3.60, Kimi K3 3.59 and GLM-5.2 3.40.
Business agents and document creation
Mistral says Large 4 can power general-purpose agents that collect information, use tools and produce completed deliverables. The company reports 59.9% on AutomationBench, which covers 657 business workflows. The applications involved include Gmail and Google Sheets, as well as Slack and Salesforce; Mistral says the model outperformed Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro in this test.
For the deliverables themselves, the company highlights the creation of spreadsheets, presentations and PDFs. It reports 1,393 Elo on AA-Briefcase, which evaluates longer-horizon knowledge work, placing the model ahead of DeepSeek V4 Pro. Mistral also says the model can edit and repair complex spreadsheets and documents.
Mistral further reports that Large 4 exceeded GPT-6-Astra in evaluations of representative legal and financial tasks conducted through vals.ai. The company says it also outperformed all open-source models on HarveyAI’s Legal Agent benchmark.
Visual analysis of drawings, PDFs and geospatial imagery
In its demonstrations, Mistral combines object recognition and localization with agentic workflows. Large 4 demonstrates identifying objects in complex scenes and checking mechanical components in technical drawings. Other demos show it retrieving supporting evidence from PDFs and searching large geospatial images, zooming in, inspecting and checking repeatedly during analysis.
The company also published a direct comparison on the Dense 200 visual grounding test. According to Mistral, Large 4 scored 42%, while GPT-6-Astra scored 41%.
Scientific coding and comparison with GLM-5.3
For scientific work, Mistral credits the model with agentic coding capabilities for data analysis, modeling and physical simulations. It claims, for example, that the model can generate a complete Hartree–Fock simulation in a single attempt. The company also describes it as the leading open-weight model on SciCode-Verified, which evaluates the implementation of complex scientific workflows in code.
Mistral’s internal evaluation also shows areas where results were more evenly matched. Expert evaluators compared Large 4 with GLM-5.3 across coding, computer-aided design, or CAD, finance, mathematics and physics. They preferred Large 4 in CAD and STEM, while its performance in finance and coding was on par with or close to GLM-5.3.
Finding vulnerabilities and resisting malicious instructions
Mistral reports 82% on a test requiring the model to reproduce a real vulnerability in open-source software and then fix it. The company describes this as the highest result of any model. On Cybench, a set of 40 exercises drawn from security competitions, Mistral says Large 4 solved 93% of the tasks.
Separately, the company reports that the model resisted 93.3% of attacks on Lakera’s B3 security benchmark. Mistral links this result to resistance against indirect prompt injection and says it has seen no higher score among competitors.
European infrastructure and preparations for the weight release
According to Mistral, Large 4 uses a MoE architecture with one trillion parameters, of which 49 billion are active. The company says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers. The same infrastructure also serves the public preview.
Mistral says a significant portion of the training data covered more than 160 languages, including every official language of the European Union. The company plans availability across multiple regions worldwide. This is set to include a European deployment operated entirely by Mistral under European law.
Ahead of the weight release, Mistral is testing the model in real-world conditions with cybersecurity experts, vetted partners and state authorities. These participants are to receive access to the same model with reduced moderation and expanded cybersecurity capabilities.
