Mistral Launches Mistral Large 4: 1-Trillion-Parameter Open-Weight AI Model
Summary
- • Mistral releases ML4, a 1T-parameter open-weight multimodal model; strongest outside China
- • Preview API live on Mistral Studio today; model weights to publicly release by end of October
- • Trained on 3,800 Grace Blackwell GPUs in Mistral's own European data centers over two months
- • Top 5 globally on AI cybersecurity benchmarks; targets finance, law, and enterprise workloads
Updates
ML4 nicknamed 'Le Chonk'; Macron calls Mistral a 'third way in AI' between US and China
Mistral's ML4 has been nicknamed 'Le Chonk' internally in reference to its 1 trillion parameters. French president Macron described Mistral as offering 'a third way in AI' — distinct from US closed-model providers and open-weight models developed primarily in China.
Chip design as optimized use case; core to ASML (Series C lead) and Samsung (Series D lead)
ML4's targeted use cases include chip design alongside cybersecurity and finance. This aligns with the interests of two key investors: Dutch semiconductor giant ASML, which led Mistral's Series C, and Samsung, which led the €21B Series D in September 2026.
Mistral faced criticism for hosting Chinese models; ML4 reasserts its frontier AI lab status
Mistral previously drew scrutiny after hosting Chinese AI models, prompting questions about whether it was pivoting toward becoming an inference provider. The ML4 launch is Mistral's most direct response — asserting its place as a frontier lab that builds its own leading models, not just serves others.
Scores 38 on AI Analysis Intelligence Index — level with GPT-6 Luna, just below DeepSeek V4.1 Flash (39)
Mistral Large 4 Preview scores 38 on the Artificial Analysis Intelligence Index, placing it level with GPT-6 Luna (max, 38) and just below DeepSeek V4.1 Flash (max, 39). Independent benchmarking confirms it is the most intelligent model developed outside the US and China.
Cyber Index score 50; CyberGym-E2E-AA at 82%, ahead of MiMo-V2.6-Pro (79%) and GPT-6 Luna (78%)
On the Artificial Analysis Cyber Index, ML4 scores 50, level with GLM-5.3-Flash (50). On the CyberGym-E2E-AA sub-benchmark specifically, ML4 scores 82%, ahead of MiMo-V2.6-Pro (79%) and GPT-6 Luna max (78%). Once weights release, it will rank among the top three open-weights models on the Cyber Index.
Pricing: $1.36/$4.18 per 1M tokens; 50% launch discount for 2 weeks; 512k context window
ML4 Preview is priced at $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14/M tokens. A 50% launch discount applies for the first two weeks ($0.68/$2.09 per 1M tokens). The model supports a 512,000 token context window.
Document reasoning: 19% on GDP.pdf (18-pt gain from ML3); API now accepts 100 images/request
ML4 Preview scores 19% on the GDP.pdf document reasoning benchmark — on par with MiMo-V2.6-Pro (19%) and behind Kimi K3 (22%). This is an 18-point improvement from Mistral Large 3, partly driven by API changes that now support 100 images per request, up from 8 for previous Mistral models.
Details
Public preview live on Mistral Studio; weights releasing end of October 2026
Mistral launched a public preview API for ML4 on Mistral Studio on October 6, 2026. Full open-weight release is planned for end of October; until then access is limited to the preview API and early partners.
1T total parameters, 49B active — sparse mixture-of-experts architecture
ML4 uses a sparse MoE design with 1 trillion total parameters but only 49 billion active at inference, enabling frontier capability with substantially lower compute per token than dense model architectures.
Natively multimodal; visual grounding surpasses some frontier closed models
ML4 is natively multimodal with exceptional performance on agentic workflows and multimodal understanding. In visual grounding tasks, it surpasses some leading closed models according to Mistral's internal evaluations.
Trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centers
ML4 was trained from scratch over two months on 3,800 NVIDIA Grace Blackwell GPUs deployed in Mistral's own European data centers. The public preview is also served on that same infrastructure, enabling full data sovereignty.
Top 5 globally on Artificial Analysis Cyber Index; leads all non-Chinese open models
On the Artificial Analysis Cyber Index — an independent benchmark for AI security and vulnerability research — ML4 ranks in the top 5 models worldwide and leads all open-weight models developed outside China by a wide margin.
Pre-release red-teaming with govts and cybersecurity leaders before public weight release
Before public weight release, ML4 is being red-teamed with cybersecurity leaders, vetted enterprise partners, and state authorities, who receive access with reduced moderation and expanded cyber capabilities for legitimate security research and incident response.
European deployment under European law; 160+ language training data including all EU languages
ML4's European deployment is operated entirely by Mistral, independently of other digital service providers and under European law. Training data covered 160+ languages, including every official EU language — a deliberate AI sovereignty design.
Launch follows €3B Series D in September 2026 at €21B valuation
The ML4 launch follows Mistral's €3 billion ($3.4B) Series D in September 2026, which valued the company at €21 billion — cementing its status as Europe's most valuable AI startup and funding the infrastructure behind ML4.
Still lags frontier closed models in some coding tasks; architecture details to follow
Mistral acknowledged ML4 still lags behind frontier closed models in certain coding benchmarks. Full architecture details, additional benchmarks, and post-training methodology will be shared ahead of the weight release.
Open-weight race: competing with Chinese GLM models and US startup Reflection AI
The open-weight AI race is intensifying globally. Chinese models (GLM 5.2, surpassed by GLM 5.3 in August 2026) dominate the field, while Nvidia-backed US startup Reflection AI unveiled its first open model Monday. ML4 positions Mistral as the leading Western challenger.
Covers ML4 launch details, architecture, infrastructure, cybersecurity benchmarks, European sovereignty design, and competitive open-weight context.
What This Means
Mistral Large 4 is Europe's most significant open-weight AI release to date — a 1-trillion-parameter model trained entirely on European infrastructure that directly challenges Chinese open models for global leadership. When weights release by end of October, ML4 will become the most capable self-hostable model available to enterprises and governments outside China, with particular strength in cybersecurity that addresses the real problem of closed-model guardrails blocking legitimate vulnerability research. The open-weight model race is now a three-way contest between US closed labs, Chinese open-source leaders, and Mistral as Europe's sovereign AI champion. ML4's release is likely to accelerate adoption of open models among European governments and enterprises with strict data sovereignty and compliance requirements.
Sentiment
Broadly excited about Europe's open-weight contender and sovereign AI angle
“That’s the part I’m most interested in: another serious model that developers can run and customize themselves. The open-weight race needs more contenders, and Europe badly needs more companies building.”
“Mistral just dropped Large 4. 1 trillion parameters, but only 49B active at a time. Their claim: best open-weight model from the US or Europe... Basically Europe saying: we’re still in this race.”
“The real buyers: enterprises and governments that want a model they fully own, with no reliance on US vendors and no Chinese weights running on their servers. Still, anyone who wrote Mistral off this year spoke too soon.”
“Mistral says the model scored 82% on a test of reproducing a real software flaw, then patching it. It says Claude Opus 5.5 and GPT-6 Astra score near zero there, because they refuse. Would you trust an open model with that work? And would you host it yourself?”
Split
~80/20 positive/neutral on open-weight sovereignty; little pushback seen on capability claims.
Sources
- Mistral unveils new AI model it says rivals best open systems from ChinaCNBC
- Mistral Says Its New AI Model ‘Le Chonk’ Is the Best Open-Weight Offering Outside of ChinaWired
- Introducing Mistral Large 4Mistral
- Meet Mistral Large 4, aka Le Chonk.X
- Mistral CEO says new AI model beats Chinese ones in some areasReuters
- Mistral Large 4Docs
- Mistral releases Mistral Large 4, dubbed "le Chonk", a 1T-parameter open-weight model for general agentic capabilities, trained on 4,000 Grace Blackwell GPUsThedeepview
- Mistral’s new 1T model aims to leapfrog closed and open rivalsTechCrunch
- Introducing Mistral Large 4: Le chonkSimonwillison
- Mistral Large 4Simonwillison
- Mistral debuts Large 4 ‘Le Chonk', a 1-trillion parameter text output model with high benchmarks planned for open weights releaseVenturebeat
- Q (@qtnx_) on X - Mistral Large 4 is still doing RL runs, keep seeing improvements (vs preview version). Release at the end of the monthX
- Artificial Analysis says Mistral Large 4 is the most intelligent model from outside the US and China, achieving results comparable to DeepSeek V4.1 Flash (max)Artificialanalysis
- Europe rejoins the fight with Chonky! Mistral Large 4 Released, Open weights end of month, who’s ready?Mistral
- Mistral drops 1T-parameter open-weight multimodal model beating closed frontier modelsX
- Mistral says "Le Chonk" can challenge the best AI modelsArs Technica
- Huge companies have dominated AI. Two startups with an entirely different approach think they can do better - CNNCnn
- Mistral Launches Preview of Mistral Large 4 "Le Chonk" 1T ModelVenturebeat