Mistral Large 4 (Le Chonk): Specs, Pricing, and Open Weights

👁 114 reads

On October 6, 2026, Mistral AI released a public preview of Mistral Large 4, known internally as "Le Chonk": a mixture-of-experts model packing 1.05 trillion total parameters, with 49 billion active per token, multimodal image input, and a 1-million-token context window. The API is live from day one at $1.36 per million input tokens and $4.18 per million output tokens (currently half price during launch), with open weights expected by the end of October. I verified the technical documentation, pricing, and immediately benchmarked the model in our evaluation suite. The comprehensive company overview is available in our Mistral AI hub.

What Mistral Announced on October 6

Mistral Large 4 is the company’s first new "Large" release since December 2025 and the biggest model in its history. Mistral describes it as a hybrid instruction and reasoning model: the exact same model can respond immediately or "think" before generating an answer, depending on your parameters. The company calls it "the best open-weights model originating from the US or Europe on aggregate benchmarks." That geographic caveat is deliberate - open-weight leaderboards have been dominated for over a year by Chinese labs: DeepSeek, Moonshot (Kimi), Alibaba (Qwen), and Z.ai (GLM).

These figures should be treated as preliminary. Initial reports circulated figures of 675B total and 41B active parameters, but those belong to its predecessor, Mistral Large 3. The exact number of experts, routing mechanism, and layer configurations will be disclosed alongside the open weights.

Mistral’s Benchmark Claims

Mistral focused heavily on enterprise and public-sector capabilities: cybersecurity, finance, law, manufacturing, and code generation.

Artificial Analysis independently evaluated the model: it scored 38.4 points on their Intelligence Index, placing 8th among open models behind seven Chinese competitors (data from Oct 6 via Trending Topics). By comparison, Claude Opus 5.5 scores 57.6 points, while the leading open model, MiMo-V2.6-Pro, reaches 46.3. For complete performance comparisons, see Mistral Large 4 vs Claude, GPT, Gemini, and DeepSeek.

Open Weights: Release Date and Licensing

• Target date: Mistral states "by the end of the month." VentureBeat reported October 27, following roughly three weeks of red-teaming with cybersecurity firms, commercial partners, and government agencies.
• Reason for the delay: Pierre Stock, VP of Science at Mistral, told TechCrunch that the team is validating with trusted partners whether the open weights can be used strictly for defensive rather than offensive cyber operations. Reports also indicate the final checkpoint is still undergoing reinforcement learning.
• License: Unconfirmed. Mistral Large 3 and Small 4 used Apache 2.0, Medium 3.5 used a modified MIT license, and VentureBeat expects a custom Mistral proprietary-open license. At this stage, Apache 2.0 cannot be assumed.

Self-hosting will be entirely out of reach for consumer hardware: 1.05 trillion parameters require roughly 2.1 TB of memory in FP16 and around 525 GB even with 4-bit quantization. This necessitates multi-node server infrastructure, not an RTX 4090. Smaller Mistral variants that run on consumer GPUs are covered in our guide to Mistral AI.

Pricing: Competitive, But Not the Cheapest

Standard list pricing of $1.36 / $4.18 per million tokens undercuts Claude Sonnet 5.5 and GPT-6.1 Sol (both $2 / $10 on OpenRouter) as well as Mistral Medium 3.5 ($1.50 / $7.50). The temporary 50% discount drops rates to $0.68 / $2.09. Mistral has not stated when the promotion ends - industry commentary suggests approximately two weeks, though this remains unconfirmed. Be mindful that reasoning mode consumes tokens rapidly: in my test suite, 42,230 of 62,423 generated tokens went to internal reasoning tokens, which are billed at standard output rates. A full pricing breakdown is available in our guide to Mistral pricing.

Where to Access Mistral Large 4

1. Mistral Studio API: Accessible via La Plateforme accounts.
2. OpenRouter: Made available on launch day, with Mistral hosting as the sole provider. I used this route for testing. Read our walkthrough on OpenRouter to get started.
3. Mistral Vibe (formerly Le Chat): Launch materials made no mention of Large 4 being deployed to the consumer chat interface yet.

Note: OpenRouter currently exposes reasoning as a binary toggle - either fully enabled or disabled - with no intermediate "low" or "medium" effort levels, which directly impacts latency and billing.

First Hands-on Evaluation

Immediately after launch, I ran Large 4 through our standardized benchmark suite (we tested it in Polish) alongside 30+ other models from our multilingual model arena: 9 text generation prompts (article, product description, customer email, social post, short story, headlines, summary, explanation for a child, translation), two proofreading challenges, and two grammar/usage quizzes. Responses were judged blindly by Claude Opus 5.5 and GPT-6 Sol.

• With reasoning enabled: 69.2 points out of 100 - placing it on par with Mistral Medium 3.5 (68.0), but trailing DeepSeek V4 Pro (74.1) and Claude Sonnet 5.5 (80.4).
• Without reasoning: 65.6 points, completing 9 texts in 91 seconds instead of nearly 15 minutes, costing just $0.015 across all 13 prompts.
• Strengths: Formal linguistic precision, scoring 46/46 and 46/48 on language quizzes.
• Weaknesses: In reasoning mode, it hit maximum token limits on three prompts during the thinking phase alone, yielding empty or truncated responses. It also occasionally invented factual claims and leaned on unnatural phrasing.

Detailed scores, test logs, and comparison charts can be found in our deep-dive review: Mistral Large 4 Tested.

The Verdict

For European AI sovereignty, this is a milestone: a massive flagship model trained from scratch inside EU data centers, supporting official European languages natively, with weights on the way. For specialized non-English text generation, it does not currently dethrone more cost-effective alternatives like DeepSeek or Gemini 3.8 Flash. Its strongest value proposition lies with enterprises seeking EU data residency compliance or organizations planning on-premises deployments once the weights drop. Just a day prior, US-based Reflection AI made a comparable play with Beam.

Sources

• Mistral: Introducing Mistral Large 4 (Oct 6, 2026)
• Model Documentation: mistral-large-4-0 (parameters, context, pricing)
• VentureBeat (weights date Oct 27, licensing)
• MarkTechPost (1.05T parameter confirmation)
• TestingCatalog (benchmarks, red teaming)
• TechCrunch (Pierre Stock statement)
• Trending Topics (Artificial Analysis Intelligence Index)
• OpenRouter model catalog and proprietary test suite, Oct 7, 2026

Verified: October 7, 2026.

// read moreMistral Large 4 Enters the AI Arena: October 2026 RankingsBenchmarksMistral Large 4 vs Claude, GPT, Gemini and DeepSeek: PricingBenchmarksMistral AI Large for GDPR-Friendly Email PersonalizationGuides

Kinetyra