Downloadable Weights: Mistral Opens Its Trillion-Parameter Large 4 for Self-Hosting
French company Mistral on Thursday 29 October made the weights of Mistral Large 4 (ML4) — nicknamed "Le Chonk" — downloadable for businesses and governments that want to run the model on their own infrastructure.
The release comes three weeks after the API preview on 6 October, and the company markets the model as a sovereign, European alternative to closed vendors and Chinese open models. But independent measurements complicate the "frontier" picture: on Artificial Analysis' cyber index, Large 4 scores 51, clearly behind its rivals — even as the open weights make the model's safeguards removable and its copies impossible to monitor, warns the UK's AI Security Institute.
What was released
Large 4 is a mixture-of-experts model with 1.05 trillion total parameters, of which 52 billion are active per token. The earlier figure of 49 billion active parameters is explained by Mistral as the larger sum also counting embeddings and output layers — an explanation that has not yet been independently confirmed (VentureBeat).
The model is multimodal (text and images as input) and, according to the company, was trained on more than 160 languages, including all of the EU's official languages (CNET). The released version has three reasoning modes — low, medium and high — so users can weigh cost against complexity, and it was delivered at the time Mistral had promised after the preview (ZDNET).
The weights can be downloaded from Hugging Face, and the model is available via Mistral Studio, Mistral API, OpenRouter, Vibe Code and Nvidia-hosted endpoints (CNET).
Mistral's narrative: sovereignty and European self-determination
The core of Mistral's argument lies with vice president of science Pierre Stock. "Concretely, it means, for example, that as a business or a government agency you get access to frontier intelligence to defend yourself against cyberattacks. Access that by definition cannot be cut off, because the model has open weights," he told TechCrunch (TechCrunch).
To Fortune, Stock called the release "a turning point for Europe" and argued that the continent can now build an ecosystem around AI foundation models, "not just the application layer." He said the model was trained on Mistral's own 10-megawatt compute cluster in France — "a fraction of what all the competitors use." Mistral has, according to Fortune, recently raised €3 billion in a Series D led by Samsung in September, at a valuation above €21 billion (Fortune via MSN).
Chief revenue officer Marjorie Janiewicz pointed to customer demand: "I don't want to be disconnected. I want to own my model," she told VentureBeat, adding: "Second, I want to be able to customize it to make sure I'm extracting value. And third, it's business continuity" (VentureBeat).
At the same time, Stock warned against Chinese open models: a model's behavior could in theory be altered to exfiltrate secrets from a company's systems, he told Fortune.
What the independent measurements show
This is where the picture becomes more layered than Mistral's marketing.
Overall intelligence. On 29 October, Artificial Analysis gave ML4 26 points on the Intelligence Index v5 in high reasoning mode, up from 25 for the preview. That makes ML4 the highest-scoring model with available weights developed outside China — but six points behind open leaders such as DeepSeek V4.1 Flash 1017 and GLM-5.3 (VentureBeat).
Cyber. On Artificial Analysis' Cyber Index — which measures how well models find and fix vulnerabilities, not carry out attacks — Large 4 scores 51. That is clearly behind Kimi K3.1 Pro (64), DeepSeek V4.1 Flash (57) and Grok 4.7 (56) (MSN/Cryptopolitan). On the defensive CyberGym benchmark, which tests finding and patching real memory-safety vulnerabilities in C/C++ projects, Le Chonk sits just above DeepSeek V4.1 Flash (ZDNET).
Coding. Mistral claims ML4 "tops" the coding index among open models, but on Artificial Analysis' aggregate Coding Agent Index, ML4 sits at 48 — one point behind GLM-5.3's 49. Codex GPT-6.1 Terra also scores 48 (ZDNET). The sources cannot be reconciled; it may be due to different sub-benchmarks or settings.
Agentic capability. Mistral says three additional weeks of reinforcement learning since 6 October lifted Terminal-Bench 4.0 accuracy from 28 to 41 percent. But Anthropic's closed Mythos 5.1 scores 60.9 percent on the same benchmark (Fortune via MSN).
The gap to the frontier. An analysis from Epoch AI on 16 October gave GLM-5.3, the leading open model, a Cyber ECI score of 160 versus 171 for the strongest closed models. Based on historical progress, Epoch estimated that open-weight models trail the closed frontier by six to seven months; AISI has estimated four to seven months (MSN/Cryptopolitan).
Economics, license — and an unsettled price
Efficiency is a strong card in Mistral's hand, but the claims must be separated from the measurements. Mistral claims ML4 achieves performance comparable to GLM-5.3 with 47 percent fewer tokens per task (TechCrunch). Independent measurement shows a 63 percent cost reduction per Intelligence Index task compared with the preview, from $5.54 to $2.06 (VentureBeat). The two efficiency figures are not directly comparable, since the underlying setups and settings are not specified in the sources.
The API price itself, however, remains unsettled in the sources. VentureBeat reports $0.70 per million input tokens and $2.10 per million output tokens. OrcaRouter, on the other hand, read Mistral's pricing page on 30 October as a list price of $1.36/$4.18 (with $0.14 for cached tokens) plus a half-price launch promotion at $0.68/$0.07/$2.09 — with no published end date (OrcaRouter). Whether the low prices are permanent or promotional remains unclear.
The license, the Mistral Ecosystem Open-Weight License, is reported as relatively open: no revenue threshold, and the requirement for a separate commercial agreement applies only if the user or an affiliated entity runs a model-as-a-service business, with certain exceptions (VentureBeat).
The security side: freedom at a cost
Mistral says Large 4 refuses 96.6 percent of malicious cyber requests — a company figure that has not been independently verified. But once the weights are downloaded, the safeguards can be removed, the reporting points out. The UK's AI Security Institute (AISI) warns that private copies cannot be monitored or recalled (MSN/Cryptopolitan). It is the same structure that makes the sovereignty argument valid: what guarantees access also guarantees that no one can shut the door.
Mistral also gave its first test partners a version with fewer safety guardrails and "extended cyber capabilities," following the pattern of Anthropic's Project Glasswing and OpenAI's Astra rollout (ZDNET). To date the risk is theoretical: as of 30 October, no documented cyberattacks have been linked to the released weights.
What this means
For Norwegian and European organizations considering on-premise AI, the release offers a real choice: a model that can be owned, customized and run behind one's own firewall, with license terms that permit most internal uses without an extra agreement. But what has been verified so far is primarily the cost per task and the overall score outside China — not leadership in cyber or agentic coding. The cyber index places Large 4 behind both Chinese and American competitors, and the gap to the closed frontier is, according to independent analyses, still measured in months, not days.
The open questions: whether the launch prices will last, whether the token savings hold up in real workloads, and whether the safeguards in the weights are actually robust against removal. Mistral's frontier claim is, as far as the sources show, a statement — not a measured result.