Mistral Large 4 (Le Chonk): Europe's 1 Trillion Parameter Model Enters Preview

On October 6, 2026, Mistral AI launched a public preview of Mistral Large 4 — a 1 trillion-parameter mixture-of-experts model nicknamed "Le Chonk" — through its Studio API. With roughly 49 billion active parameters per token, a 1-million-token context window, and native multimodal training, Large 4 is Mistral's bid for the European sovereign open-weight frontier. The preview runs on Mistral's European data centers; open weights are promised by end of October 2026.

What Mistral actually shipped

Mistral's October 6 announcement opens with a simple line: "Today, we're launching a public preview of Mistral Large 4." The preview runs through the Mistral Studio API under model ID mistral-large-4, giving developers a way to test before any weights become downloadable. Mistral says Large 4 was the first milestone funded by its €3 billion Series D — the largest equity round ever raised by a European technology company.

The core spec sheet:

• Architecture: Mixture-of-Experts (MoE), natively multimodal (text + images from pretraining)
• Total parameters: 1 trillion
• Active parameters: 49 billion per token (52B including embeddings/output layers)
• Context window: 1 million tokens (Artificial Analysis lists 512K; treat as provisional until model card)
• Languages: 160+, including every official EU language
• Training: ~3,800 NVIDIA Grace Blackwell GPUs in Mistral-operated European data centers
• Pricing (preview): $1.36/M input, $4.18/M output; cached input $0.14/M

Why the parameter count matters — and why it was controversial

A trillion total parameters with only 49 billion active tells you this is a sparse MoE: each token routes through a small subset of expert sub-networks, storing a trillion parameters' worth of knowledge but paying the compute cost of roughly a 49B dense model at inference. The 20:1 total-to-active ratio is notably sparse even by current MoE standards.

Earlier reports circulated figures of 675 billion total / 41 billion active (from Mistral Large 3) and 1.05 trillion / 52 billion. Mistral's official post confirms 1 trillion / 49 billion; treat those as authoritative going forward.

Access, pricing, and what "preview" means

The preview is available today on Mistral Studio. A typical request:

curl https://api.mistral.ai/v1/chat/completions       -H "Authorization: Bearer $MISTRAL_API_KEY"       -H "Content-Type: application/json"       -d '{"model": "mistral-large-4", "messages": [{"role": "user", "content": "Summarize this quarterly report."}]}'
    

Caveats: This is a preview, not a stable release. Expect rate limits, possible latency spikes, and behavior changes before final weights ship. Mistral's pricing page also lists an example of $0.50/$1.50 for "Mistral Large" — that applies to the older model, not Large 4. Use the $1.36/$4.18 rates for budgeting.

Mistral aims the model at coding, agentic, and cybersecurity workloads. The announcement highlights work with NVIDIA on engineering/industrial AI workflows and describes Large 4 as trained on Grace Blackwell GPUs in European data centers — a signal about the target buyer.

Benchmark claims: vendor-reported vs independent

Mistral reports (self-reported, from announcement):

• Cybench: 93%
• Reproduce-and-patch vulnerability test: 82%
• DeepSWE v1.1: 61.7%
• Lakera B3 prompt-injection resistance: 93.3%
• Coding Agent Index (Artificial Analysis): 49.8%
• AutomationBench: 59.9%
• CyberGym-E2E: 82%
• Blind coding evaluation (with Surge AI): ranked 2nd of 5, behind Claude Opus 5.5

Artificial Analysis (independent, read October 21, 2026):

• Intelligence Index: 38 (below frontier tier)
• Cyber Index: 50
• CyberGym-E2E-AA: 82%
• Cost per task: over 4x that of similar-intelligence open models

The divergence is significant. Mistral's own numbers position Large 4 as a cybersecurity specialist with frontier-level coding; Artificial Analysis sees a model with moderate general intelligence and high cost-per-task. No third-party replication exists yet. Teams evaluating should run their own tests on their own documents.

European sovereignty: the practical angle

Mistral's announcement is specific about the sovereignty angle: the preview runs on the same European infrastructure used for training, a European deployment option is operated end-to-end under European law, and organizations can self-deploy on private cloud or on-premise hardware. For security operations, banks, public agencies, and defense suppliers that cannot send logs, malware samples, or contracts to US-hosted APIs, this matters.

The EU's AI Act places obligations on general-purpose model providers, pushing procurement teams toward vendors with a clear European legal footprint. A strong model that runs inside your own perimeter removes a whole category of compliance review.

Open weights: what to watch

Mistral promises open weights "by the end of October 2026" (Hugging Face placeholder: October 31). Several unknowns remain:

• Exact license terms (Large 3 is Apache 2.0; Large 4 docs list only "Open")
• Number of experts, routing topology, shared experts
• Attention design, tokenizer, vision encoder
• Verified long-context performance (no published needle-in-haystack results)

Self-hosting at FP8 needs ~1 TB of weights — a memory commitment, not just compute. Check the model card on release day before planning any production deployment.

Bottom line: Mistral Large 4 preview is live today at $1.36/$4.18 with a 1M context window and native multimodal support. The 1T/49B MoE architecture is compute-efficient at inference but memory-heavy for self-hosting. Cybersecurity benchmarks are the standout claim; general intelligence scores are contested. Open weights promised by end of October — check the license before building on it.

Prices and benchmark figures are standard published rates as of October 24, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.