Mistral Large 4 'Le Chonk': What a 1-Trillion-Parameter Open Model Means for Your Chatbot
Mistral just released Le Chonk, a 1T open-weight model trained in European datacenters. Here is what it changes for your business chatbot in 2026.
On October 6, 2026, Mistral AI quietly announced the single biggest shift in European enterprise AI since the EU AI Act: Mistral Large 4, nicknamed “Le Chonk”, a 1.05-trillion-parameter mixture-of-experts model with the open weights released on October 27. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters, and on independent benchmarks it is now the top open-weight model built outside China.
If you run a business chatbot in Europe, this is not a product launch to scroll past. It is the moment a frontier model you can actually audit, host sovereignly, and deploy at predictable prices became the default option. Here is what Le Chonk means, concretely, for the chatbot on your website.
What is actually inside Mistral Large 4
Le Chonk is a granular mixture-of-experts architecture: 1.05 trillion total parameters, only 49 billion active per token, a 1.6-billion-parameter vision encoder, and a 1 million token context window. The preview API is live on Mistral Studio at $1.36 per million input tokens and $4.18 per million output tokens — roughly a quarter of what you pay for GPT-6 Astra or Claude Fable 5.1 at equivalent capability.
The benchmark scores are, politely, embarrassing for the US closed-weight incumbents:
- Cybench: 93% and CyberGym-E2E: 82% on vulnerability reproduction and patching
- Lakera B3: 93.3% of prompt-injection attacks resisted
- Harvey Legal Agent: #1 among all open-source models
- Dense 200 visual grounding: 42% — narrowly above GPT-6 Astra at 41%
- DeepSWE v1.1: 61.7% on autonomous software engineering tasks
Three details matter more than the raw numbers. First, the weights ship on October 27 under an open-weight license — you can inspect them, fine-tune them, and if required host them inside your own perimeter. Second, the training compute is European-sited, so there is no US re-export question hanging over the stack. Third, Mistral is deliberately positioning the model as “the top open-weight model from the West,” a message that lands differently depending on whether you are a European CTO or a US procurement team.
Why this is a strategic shift for European chatbots
For eighteen months, the trade-off facing European businesses was uncomfortable: either rent a US closed-weight model (ChatGPT, Claude, Gemini) and inherit the CLOUD Act, pay the EU AI Act Article 50 audit tax on top, and watch your competitor’s data cross the Atlantic in both directions; or settle for a smaller open model that was noticeably behind on reasoning, coding, and multimodal tasks.
Le Chonk ends that trade-off. The same month that CNIL, BfDI and AESIA opened the first wave of EU AI Act inspections, European teams can now point to a frontier-class open-weight model that beats several closed US systems on security and legal reasoning, is cheaper per token, runs in European datacenters by default, and is governed by European law.
The downstream implication for business chatbots is simple. The “quality of answers” argument that US vendors used to charge a premium for sovereignty is gone. A RAG chatbot sitting on Le Chonk through Scaleway is now objectively more secure (93.3% on Lakera’s prompt-injection benchmark is the highest ever recorded for an open model), objectively cheaper, and structurally compliant with GDPR, the DSA, and EU AI Act Article 50. The only remaining excuse for keeping customer conversations on a US stack is inertia.
Cybersecurity leadership is a chatbot story, not an infrastructure story
The headline Cybench and CyberGym-E2E scores read like cybersecurity news, but the number that actually matters for a chatbot on your website is 93.3% on Lakera B3. Lakera B3 is a prompt-injection benchmark: it measures how well a model refuses to leak system prompts, bypass safety instructions, or follow adversarial content hidden inside documents that get ingested into a RAG pipeline.
Prompt injection is the OWASP top risk for LLM applications, and 55% of real-world attacks on RAG chatbots are “indirect” injections embedded in documents the business itself indexed — a PDF with a hidden instruction, a scraped web page with an invisible paragraph, a Markdown note smuggling a <prompt> tag. For a sovereign chatbot stack like DoxyChat, where visitor-facing bots are fed by whatever files the customer uploads, “the model refuses the injection” is a structural safety guarantee, not a nice-to-have.
Mistral already shipped Shieldstral in August 2026 as a standalone 3-billion-parameter safety classifier. Le Chonk now folds that philosophy directly into the base model. For a RAG chatbot operator, this is one less third-party moderation layer to buy, maintain, and prove to the CNIL.
What this changes for DoxyChat customers
DoxyChat has used Mistral as its primary LLM since day one, served from Scaleway’s French datacenters with Gemini as a fallback. That architectural decision — made months before most of our competitors even acknowledged sovereign AI as a category — now pays off automatically.
The moment Mistral Large 4 lands in Scaleway’s inference catalogue, every chatbot built on DoxyChat gets:
- Better reasoning on complex questions. A million-token context window means the retrieval pipeline can hand the model more candidate chunks without truncation, which directly improves answer precision on long technical documents (legal contracts, insurance policies, product manuals, HR handbooks).
- Lower per-conversation cost. $1.36 / $4.18 per million tokens is a 60–75% price cut versus the proprietary APIs most “chatbot platforms” quietly bill their customers for.
- Measurably stronger prompt-injection resistance. The same documents that previously required extra moderation passes can now be ingested with less overhead, because the base model itself declines the most common attacks.
- No contract renegotiation when the US changes its mind. Open weights mean the model does not disappear if an export control, a Section 301 tariff, or a Supreme Court ruling changes the rules — the exact scenario European banks hit when Anthropic was briefly export-restricted in June 2026.
None of this requires a migration, a plan upgrade, or a conversation with your legal team. The stack was already pointing in this direction.
The honest limitation
Le Chonk’s open weights do not ship until October 27, and the Scaleway integration will land a few weeks after that. In the meantime, DoxyChat continues to serve on Mistral Medium 3.5 — still a European-hosted, GDPR-native model that outperforms most US generalist chatbots on documentary QA. The upgrade will be invisible to end users: same widget, same URL, same 1-line JavaScript integration.
The decision that was always coming
A year ago, choosing a sovereign AI chatbot meant accepting a measurable quality gap for a measurable compliance win. On October 6, 2026, that gap closed. For every French PME, every German Mittelstand firm, and every Spanish or Italian business that has been waiting for “the European model to catch up” before committing — the waiting is over.
If you are still running your customer-facing chatbot on a US closed-weight API, the question is no longer “is European AI good enough yet?”. It is “why am I still paying 4x for a less secure answer?”
Try DoxyChat free at www.doxychat.com — the Discovery plan gives you one chatbot, 10 documents, and 200 requests per month on the same Mistral + Scaleway stack that will host Le Chonk at the end of the month. No credit card, no US data transfer, no Article 50 audit paperwork to retrofit later.