--- title: Europe Has the Top Open-Weight AI Model Outside China With Mistral Large 4 description: Mistral Large 4 scores 26 on the Artificial Analysis index, top among open models outside China, and comes within a point of GLM-5.3 on coding. author: Darie Nani (Editor-in-Chief) date: 2026-10-30T18:39:39.726Z updated: 2026-10-30T18:39:39.735Z canonical: https://www.sovereignmagazine.com/article/mistral-large-4-open-weight-benchmarks image: https://cdn.nanimediahouse.com/le-chonk-mistral-large-4.webp categories: Artificial Intelligence content_type: News region: France publication: Sovereign Magazine about: - type: Organization name: Mistral AI --- Mistral released the weights of its 1-trillion-parameter Large 4 model on 29 October, giving the Paris lab the highest-scoring open model from outside China on the Intelligence Index run by Artificial Analysis, an independent benchmarking firm. Large 4 scores 26 on the index, ahead of Muse Spark Open from Meta on 24.7. Coding is the work Mistral chose to compete on, and there the model scores 48.3 on the Artificial Analysis Coding Agent Index, against 48.9 for GLM-5.3 from China's Z.ai, the highest score of any open model. In June we reported that [Chinese open-weight labs ranked above Europe's best on public leaderboards](https://www.sovereignmagazine.com/article/europe-sovereign-ai-talk-vs-action). On the overall index they still do, with models from six Chinese labs scoring higher than Large 4. On coding, the gap has closed to less than a point. ## Open-Weight Models Outside China Meta, the closest rival, released Muse Spark Open, an 816-billion-parameter model, this month alongside its closed Muse Wonder models, two months after [the much smaller Muse Glimmer](https://www.sovereignmagazine.com/article/meta-muse-glimmer-open-30b-apache-2). Beam 1.0, the first open model from the New York start-up Reflection AI, and Nvidia's Nemotron 3.5 Super VL score well below both. Gemma 4 from Google, Motif 3 from South Korea and Kimi K3 from China's Moonshot AI have no score on the current version of the index. ![Bar chart of the Artificial Analysis Intelligence Index showing Mistral Large 4 on 26.0, the highest open model outside China](https://cdn.nanimediahouse.com/mistral-large-4-intelligence-chart.webp) Meta's licence allows commercial use for up to $20 million of revenue and requires a "Built with Muse Spark Open" notice. Mistral's, a modified MIT licence, lets anyone use, modify and sell the model, while businesses that offer it as a hosted service need a separate commercial agreement. Large 3, the previous flagship, was released under Apache 2.0 with no such condition. ## Mistral Large 4 Benchmarks Large 3, released last December, scored 9 on the previous version of the Intelligence Index. The Large 4 preview scored 38 on that scale in early October, more than four times as much, and Artificial Analysis called it the most intelligent model from outside the US and China. Artificial Analysis has since moved to a new version of the index, so the final model's 26 cannot be set against those figures. Mistral spent the three weeks after the preview on further reinforcement learning, which trains a model by rewarding it for completing tasks. Terminal-Bench 4.0, a test of agent work at the command line, rose from a self-reported 28% for the preview to 41.4% for the final model, and Mistral says the model used 30% fewer output tokens to get there. That result puts Large 4 second among open models on the Coding Agent Index. Across the 34 pairings of agent and model on the index it ranks 15th, above OpenAI's GPT-6.1 Terra and Grok 4.7. It also uses 8.5 million tokens per task against 15.9 million for GLM-5.3, or 47% fewer. ![Bar chart of the Artificial Analysis Coding Agent Index showing Mistral Large 4 on 48.3, behind GLM-5.3 on 48.9](https://cdn.nanimediahouse.com/mistral-large-4-coding-chart.webp) The Coding Agent Index tests each model inside a coding agent, Mistral's own Vibe in this case, and the 41.4% comes from that run. The run Artificial Analysis uses for its Intelligence Index records 24.7% for Large 4 on the same test, behind GLM-5.3 on 41.9%. Mistral says Large 4 also leads the FrontierCode 2.0 test from Cognition with 19.7%, while the Artificial Analysis run of that test puts DeepSeek V4.1 Flash 1017 on 28.5% and Large 4 on 15%. ## Where China's Open-Weight Labs Stay Ahead Mistral's model card says the leading Chinese open models score higher on the Intelligence Index and that Large 4 is particularly strong on coding and cybersecurity. On the Artificial Analysis Cyber Index it scores 51, behind DeepSeek V4.1 Flash 1017 and MiMo-V2.6-Pro from Xiaomi, and Mistral's own charts show Kimi K3 ahead on FinWorkBench, a finance test, and BrowseComp, a web research test. Large 4 lists at $0.70 per million input tokens and $2.10 per million output tokens. Artificial Analysis puts the average cost of a task on its index at $2.06, more than three times the cost for DeepSeek V4.1 Flash 1017 and less than half the cost for GLM-5.3. ## Mistral's Sovereign AI Pitch to Banks and Security Agencies Mistral says it trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres, and it describes the model as having clear lineage from inception to release. Five of the six labs that outscore it have been accused of training their models on the outputs of American rivals, a practice known as distillation: [US security agencies named DeepSeek, Z.ai, MiniMax and StepFun in a September advisory, and Anthropic has named Xiaomi](https://www.sovereignmagazine.com/article/kimi-deepseek-claude-prompts-anthropic). China's Commerce Ministry calls the accusations groundless. No rival lab or government has accused Mistral of the same. The nearest comparison is a claim in August 2025 by a former employee that an earlier Mistral model had been distilled from DeepSeek, which was never confirmed. ENISA, the EU's cybersecurity agency, says it is testing Large 4 to find security vulnerabilities, audit code and analyse defensive logs. Orange Cyberdefense says it is building European cyber AI with Mistral, and ABN AMRO, Belfius, Crédit Agricole and La Banque Postale gave statements in support of the release. Investors have backed the same pitch, with Mistral raising €3 billion in September at a valuation above €21 billion in a round led by Samsung Electronics. Mistral reports that Large 4 refuses 96.6% of malicious cyber requests, the highest rate it has measured among open models. It also says its monitoring of live traffic since the preview showed "a potential for uplift to conduct real attacks", and it is changing its guardrails as a result. A specialised cyber variant is in preview with partners such as Orange Cyberdefense. Large 4 was trained on American chips, and Microsoft is the first cloud provider to offer it. Self-hosting also takes serious hardware, with Mistral's Hugging Face examples running on eight GPUs. ## The Gap to Closed Frontier Models > "It's certainly not a Fable-class model, but it's great to see Mistral put out a model that's back to being maybe about 6 months behind the frontier." > — Simon Willison Willison, a developer and writer, was assessing the preview on 6 October. Gemini 4 Carbon from Google and Claude Opus 5.5 from Anthropic now lead the Intelligence Index, more than 30 points above Large 4, and the distance on coding is about 13 points. In a blind comparison shown on Mistral's own chart, human graders preferred the answer from Opus 5.5 on 78% of prompts. They preferred Large 4 on 64% of prompts against GLM-5.3 and against Kimi K3. Mistral says the gains from reinforcement learning "continue to accelerate, with no signs of saturation", and that a full technical report is coming. Find out more at [mistral.ai](https://mistral.ai). ## FAQ **Q: What's the best open-weight model?** DeepSeek V4.1 Flash 1017 and GLM-5.3 lead the Artificial Analysis Intelligence Index among open models, on 32.5 and 31.7. Mistral Large 4 is the highest-scoring open model from outside China on 26, and on the Coding Agent Index it is within a point of GLM-5.3. **Q: Is Mistral AI as good as ChatGPT?** Not on the benchmarks in this article. GPT-6 Astra from OpenAI scores 56.9 on the Intelligence Index against 26 for Large 4, and Codex running GPT-6 Astra scores 54.7 on the Coding Agent Index against 48.3. Large 4 can be downloaded and run on a company's own hardware, which is not possible with GPT-6 Astra. **Q: Are open-weight models free?** The weights of Large 4 can be downloaded at no cost. Companies then pay for their own hardware, and Mistral's examples use eight GPUs, or they pay Mistral $0.70 per million input tokens and $2.10 per million output tokens through its API. Businesses that offer the model as a hosted service also need a commercial licence.