Mistral Releases Le Chonk: A Trillion-Parameter Open-Weight Model Aimed at Cybersecurity and Sovereign AI

French AI company Mistral has released the weights for Mistral Large 4, a 1-trillion-parameter mixture-of-experts model nicknamed "Le Chonk," making it available for download on Hugging Face and through its API. The release follows a 23-day public preview and three additional weeks of reinforcement learning that the company says drove significant gains in coding, cybersecurity, and token efficiency.

What Was Released

Mistral Large 4 is a natively multimodal MoE model with 1.05 trillion total parameters and 49 billion active parameters per token (52 billion including embeddings and output layers). It supports a 1-million-token context window and 160-plus languages, including all official EU languages. The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters [1][10].

The company provides three reasoning modes—low, medium, and high (default)—letting developers trade latency and token usage for depth of reasoning on a single model [1][10]. A quantized NVFP4 checkpoint targets Blackwell GPUs with vLLM and SGLang support [10].

Weights are released under the Mistral Ecosystem Open-Weight License, a modified MIT license that permits commercial use, modification, and distribution but requires a separate commercial agreement for entities operating model-as-a-service businesses [1][5]. API pricing is set at $0.70 per million input tokens, $0.07 for cached input, and $2.10 per million output tokens [1].

Coding and Agentic Performance

Mistral claims Large 4 leads open-weight models on agentic coding benchmarks. On the Artificial Analysis Coding Agent Index, it scores 48% at high reasoning—effectively tied with GLM-5.3 at 49%—while using 47% fewer tokens per task (8.5 million vs. 15.9 million) [1][5]. On Terminal-Bench 4.0, the score rose from 28% at preview to 41%, edging past GLM-5.3 at 40% [1][2]. Independent evaluator Artificial Analysis, however, reports ML4 at 25% on Terminal-Bench in its separate Intelligence Index evaluation [5]. On Cognition's private FrontierCode 2.0 benchmark, which measures whether a repository maintainer would merge an agent's change, Mistral reports 19.7%, ahead of GLM-5.3 (17.8%) and DeepSeek V4.1 Flash (19.3%) [1].

Artificial Analysis confirms the Coding Agent Index ranking but notes the one-point gap is not statistically significant. Its separate Intelligence Index evaluation shows a different FrontierCode result: 15% for ML4 versus 29% for DeepSeek and 22% for GLM-5.3 [5]. Mistral also cites a Surge Code blind preference study where annotators favored ML4 over GLM-5.3 (64%), Kimi K3 (64%), and DeepSeek V4.1 Flash (56%), though it won only 22% of comparisons against Claude Opus 5.5 [1].

Cybersecurity Focus

The release is explicitly pitched at defensive cybersecurity. On the Artificial Analysis Cyber Index, ML4 scores 51% at high reasoning, placing third among current open-weight models behind DeepSeek V4.1 Flash 1017 (57%) and MiMo-V2.6-Pro (56%) [5][6]. On CyberGym-E2E—a test of finding, reproducing, and patching real memory-safety bugs in C/C++ projects—Mistral reports 79%, matching MiMo and beating DeepSeek at 76% [2]. The company also cites an 82% score on an end-to-end vulnerability reproduction and patching test, and a 96.6% average refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm [1][6].

Mistral acknowledges that open weights mean safeguards can be stripped via fine-tuning. Pierre Stock, VP of Science, told Reuters: "it's impossible to contain access" [6]. The company says it is modifying guardrails after observing potential for uplift to conduct real attacks in live traffic, and is providing a specialized cyber variant to vetted enterprise partners (including Orange Cyberdefense) under contractual safeguards for penetration testing and red-teaming on authorized systems only [1].

Efficiency and Cost

Artificial Analysis reports ML4's cost per Intelligence Index task fell 63% from preview ($5.54) to release ($2.06), driven by lower token consumption and reduced API rates [5]. However, DeepSeek V4.1 Flash 1017 ($0.60) and MiMo ($0.80) achieve higher Intelligence Index scores (32 and 30 vs. ML4's 26) at lower measured task costs [5]. ML4's hallucination rate on AA-Omniscience dropped from 41% to 28%, matching Muse Spark Open, but its factual accuracy remains 26% versus 40% for DeepSeek and 42% for Muse [5].

Sovereignty Narrative

Mistral frames the release as a sovereign AI play. CEO Arthur Mensch told TechCrunch that AI should "be a tool for empowerment, not dominance." [3]. The company emphasizes European training infrastructure, data residency, and the inability to cut off access once weights are distributed—contrasting with the June 2026 U.S. restriction on Anthropic model access for foreign nationals [8]. Enterprise testimonials from La Banque Postale, ABN AMRO, Belfius, Crédit Agricole, ASML, and ESA cite sovereignty, data control, and regulatory compliance as drivers [1][8].

What Remains Unverified

• The 82% vulnerability reproduction score and 96.6% refusal rate are vendor-reported without independent replication [6].
• API pricing for the preview period differed ($1.36/$4.18 per million); the new rates apply to the released checkpoint [5].
• No independent long-context accuracy curve has been published for the 1M window [6].
• Business outcomes from enterprise deployments are not yet quantified; Marjorie Janiewicz, CRO, said systems incorporating ML4 are still being built [5].
• The specialized cyber variant for partners remains in preview with no public benchmark data [1].

Competitive Context

ML4 is the highest-scoring Western-developed open-weight model on Artificial Analysis Intelligence Index v5 (26 points), but seven Chinese open models score higher, led by DeepSeek and GLM-5.3 at 32 [5]. Mistral reports roughly $6 billion raised over three years versus "hundreds of billions" for U.S. frontier labs [1][4]. The model activates 52B parameters on 3,800 GPUs (approximately 10 megawatts), which Stock estimates is 3–4x less compute than Chinese competitors and orders of magnitude less than U.S. closed labs [4].

The release adds a credible trillion-parameter option to the open-weight landscape, particularly for organizations prioritizing self-hosting, European data sovereignty, and defensive cybersecurity workflows. Independent benchmarking and production pilots over the coming months will determine whether the efficiency gains and capability claims translate into measurable operational advantage.