Navigate this guide · 2,075 words
The essential signal
- Mistral published the weights of Large 4 on 29 October 2026. The model holds about 1 trillion parameters with 49 billion active per token and a 1 million token context window. Anyone can download it from Hugging Face for self-hosting.
- Mistral's own scores lead open models on coding. Independent assessment by Artificial Analysis ranks it second on coding and third on cyber defence. The larger gain sits in efficiency and cost rather than raw capability.
- The licence is a modified MIT licence. Use and modification and fine-tuning and commercialization are free. Businesses that operate a model-as-a-service offering need a separate commercial licence from Mistral first. The licence sets no revenue threshold.
What Mistral released on 29 October
Mistral published the weights of Mistral Large 4 on 29 October 2026. The model carries the nickname Le Chonk. It had spent three weeks as a public preview through Mistral's API after its 6 October introduction. The weights release lets any organisation download the model and run it on its own systems instead of calling Mistral's hosted service. [1][3][6][1][3][6][7][8][10]
The download sits on Hugging Face under the name Mistral-Large-4-1T-A52B. The same model remains available through Mistral Studio and the Mistral API and through OpenRouter and NVIDIA-hosted endpoints. Microsoft Foundry is the first cloud platform to carry it with private deployment options for Azure developers. [1][7][8][1][3][6][7][8][10]
Mistral says the released version is not the preview build. Three further weeks of reinforcement learning plus feedback from customers and security partners improved coding and defensive cybersecurity and reasoning efficiency. The company describes the gains as still accelerating with no signs of saturation. [1][6][10][1][3][6][7][8][10]
What the model is
Large 4 is a mixture-of-experts model with about 1 trillion parameters in total. Only a fraction activates for each token. Hugging Face lists 49 billion active parameters per token. Mistral's release materials cite 52 billion. The two figures describe the same model. The smaller count covers activity per token while the larger count also includes embeddings and output layers. [5][7][11][5][7][9][11]
Each token routes to four of 128 experts plus one shared expert that always runs. The text decoder has 72 layers. Its attention system pairs multi-head latent attention with a sparse attention indexer that selects up to 2,048 earlier positions per query. A 1.6 billion parameter Pixtral vision encoder reads images. The model accepts text and image input and produces text and code. [5][7][5][7][9][11]
The context window holds 1 million tokens in a single pass. That is enough room for very large codebases or document sets. Mistral trained the model from scratch on 3,800 NVIDIA Grace Blackwell graphics processors in its own datacenters in Europe. It reports fluency in more than 160 languages including every official language of the European Union. [5][7][9][5][7][9][11]
The coding scores Mistral claims
Mistral calls Large 4 the best open-weight model for coding. Its headline evidence is Cognition's private FrontierCode 2.0 benchmark. That test asks whether a repository maintainer would accept an agent's code change. Mistral scores 19.7 percent and leads the open models Mistral lists. DeepSeek V4.1 Flash 1017 scores 19.3 percent and GLM-5.3 scores 17.8 percent on the same chart. [1][6][11][1][6][7][11]
On the Artificial Analysis Coding Agent Index the model scores 48 percent. Mistral describes that result as effectively tied with GLM-5.3 as the leading open-weight model. The efficiency claim sits beside the score. Large 4 uses 47 percent fewer tokens per task on average than GLM-5.3 in Mistral's comparison. [1][6][7][1][6][7][11]
Terminal-Bench 4.0 moved from 28 percent at preview to 41 percent in the weights release. Mistral says output token use fell 30 percent over the same period. Blind human preference tests run by Surge Code preferred Large 4 over GLM-5.3 and Kimi K3 and DeepSeek V4.1 Flash. Only Claude Opus 5.5 ranked ahead of it among the models tested. [1][6][7][1][6][7][11]
What independent testing says instead
Independent assessment tells a narrower story. Artificial Analysis tested the released checkpoint at its high reasoning setting and scored it 26 on Intelligence Index v5. The preview scored 25 on the same scale. That makes Large 4 the highest-scoring open-weight model developed outside China in that evaluation. It sits six points behind DeepSeek V4.1 Flash 1017 and GLM-5.3 at 32. [4][12][14][4][12][14]
On the Coding Agent Index the independent ranking puts Large 4 second rather than tied first. The tested configurations scored 49 for Opencode with GLM-5.3 at maximum reasoning and 48 for Mistral Vibe with Large 4 at high reasoning. The token gap held up. Large 4 used 8.5 million tokens per task against 15.9 million for GLM-5.3. A one-point gap alone does not establish statistical significance. [4][12][14][4][12][14]
Two of Mistral's own figures also come out lower in the independent context. Artificial Analysis reports 25 percent on Terminal-Bench 4.0 against Mistral's 41 percent and 15 percent on FrontierCode 2.0 against Mistral's 19.7 percent. The figures come from different evaluation contexts and agent configurations. Neither result should be swapped for the other or presented as universal performance. [4][12][4][12][14]
The cyber pitch
Cybersecurity is the second half of Mistral's pitch. The model scores 51 percent on the Artificial Analysis Cyber Index. That suite measures whether a model can find a flaw in a codebase and reproduce and validate it and patch it without breaking existing functions. Mistral says the model handles legitimate defensive work without unnecessary refusals. [1][6][8][1][5][6][7][8]
One sub-test stands out. The reproduce-and-patch task asks a model to recreate a real vulnerability in open-source software and then fix it. Large 4 scores 82 percent on that test. Mistral says several leading closed models score near zero on the same test because they refuse to attempt the work. Defending software often starts with proving a flaw is real. [5][7][1][5][6][7][8]
The model also solves 93 percent of Cybench. That set holds 40 challenges drawn from security competitions. Guardrails distinguish defensive tasks from malicious ones in Mistral's design. The model refuses an average of 96.6 percent of malicious cyber requests across JailbreakBench and StrongREJECT and AgentHarm. Mistral describes that refusal rate as the highest it measures among open models. [1][6][8][1][5][6][7][8]
The cyber ranking gap
The cyber headline needs the same care as the coding headline. On the Cyber Index Mistral showcases, Large 4 ranks third among current open-weight models. DeepSeek V4.1 Flash 1017 leads at 57 percent. Xiaomi's MiMo-V2.6-Pro follows at 56 percent. Mistral's 51 percent trails both. The lead Mistral claims for defensive security does not survive that chart. [11][12][14][1][6][8][11][12][13][14]
The contrast with closed models cuts the other way. Google's Gemini 4 Carbon arrived the same week as a cybersecurity-focused frontier model. It is gated to selected defenders through a Fairwind program while Google runs voluntary government testing. Large 4 is downloadable by anyone. The two releases frame opposite answers to the same question about who should hold powerful security tools. [8][13][1][6][8][11][12][13][14]
Price sharpens the contrast. Mistral charges $0.70 per million input tokens and $2.10 per million output tokens through its API. That is roughly half the $1.36 and $4.18 charged at preview. A security team that can self-host Large 4 avoids sending source code to a third-party API entirely. That control argument is the core of Mistral's case. [1][6][13][1][6][8][11][12][13][14]
The licence condition
The weights ship under the Mistral Ecosystem Open-Weight License. It is a modified MIT licence. It permits free use and modification and fine-tuning and commercialization. One condition changes the picture. A business that operates a model-as-a-service offering must obtain a separate commercial licence from Mistral before commercial use. [1][7][9][1][4][7][9][12][15]
The condition turns on what a business operates rather than what it earns. The licence contains no revenue threshold. That answers a question left open before the release. Earlier speculation asked whether Mistral would repeat a revenue gate from its existing modified MIT terms. The published file imposes no such gate. [4][15][1][4][7][9][12][15]
Two exclusions narrow the condition. End-user products that embed model capability inside specific features fall outside the model-as-a-service definition. Services that merely relay requests to models hosted by others fall outside it too. Internal use and access through Mistral's official products are exempt. The announcement does not define where wrapping a model into a software service ends and operating a model service begins. [4][9][12][1][4][7][9][12][15]
The dual-use warning
Mistral is unusually direct about the risk its own release creates. The company monitored cybersecurity-related use in live traffic after the preview. It observed the model refusing black-box attacks against live targets. It also wrote that it sees a potential for uplift to conduct real attacks. Guardrails are being modified in response with further updates promised over coming months. [1][6][8][1][3][6][7][8][12][13]
Open weights cannot be recalled. Anyone can download them. A determined user can try to fine-tune refusals away. Mistral's vice president of science Pierre Stock put the philosophy plainly to Reuters. He said he is not comfortable living in a world in which open weights do not exist. Once the model is out, he added, it is impossible to contain access. [3][13][1][3][6][7][8][12][13]
The strongest version of the model stays restricted. A specialized variant tuned for penetration testing and vulnerability assessment and red-teaming goes only to vetted enterprise partners. Orange Cyberdefense is named among them. Access sits under contract with testing restricted to owned or explicitly authorized systems in controlled environments. [1][7][12][1][3][6][7][8][12][13]
Reasoning modes and cost
Three reasoning modes let developers trade performance against cost. Low suits simple tasks and replaces the previous None mode. Medium covers everyday work. High is the default for the hardest requests that need thorough reasoning. The choice changes results as well as bills. [1][6][12][1][4][6][12][14]
Artificial Analysis measured Medium at 35 points on its Coding Agent Index at $0.78 per task attempt. High scored 48 points at $1.26 per attempt. Teams can reserve deeper reasoning for difficult debugging and use a lower setting for routine requests. Whether that saves money depends on how often a cheaper attempt needs correction. [12][4][1][4][6][12][14]
The API price fell with the weights release. Input costs $0.70 per million tokens and cached input costs $0.07. Output costs $2.10 per million tokens. Artificial Analysis puts the cost per Intelligence Index task at $2.06 against $5.54 for the preview. That is a 63 percent reduction. Chinese rivals still bill less per task. DeepSeek costs $0.60 and MiMo costs $0.26 on the same measure. [4][6][14][1][4][6][12][14]
Sovereignty: who this is for
Mistral frames the release as sovereignty rather than scale. Sovereignty means an organisation keeps control of the models it depends on and the intelligence it creates. A downloaded model cannot be withdrawn by a provider changing policy. A team can inspect it and customize it around proprietary data and run it under its own rules. [1][3][1][3][4][8][10]
Pierre Stock gave the argument its cleanest form to TechCrunch. He said access to frontier intelligence for self-defense cannot be cut off by definition because the model is open-weighted. A business or a government can deploy it on premises inside its own ecosystem. That matters most where code and logs cannot leave the building. [3][8][1][3][4][8][10]
Customers describe the same demand in plainer terms. Mistral's chief revenue officer Marjorie Janiewicz told VentureBeat that customers want to own the model and customize it and keep running it regardless of provider decisions. Named testers include Booz Allen and CrowdStrike and Elastic and Europe's cybersecurity agency ENISA. Banks including ABN AMRO and La Banque Postale frame sovereignty as a condition of use rather than a feature. [4][10][1][3][4][8][10]
What self-hosting actually takes
Downloading weights is not the same as running them cheaply. The official checkpoints ship in FP8 and NVFP4 formats. Community quantizations in GGUF format come from Unsloth and the llama.cpp team. Serving tools with day-one support include vLLM and SGLang and Transformers and llama.cpp. An embedded DSpark speculative-decoding draft head proposes blocks of three tokens. Mistral measured up to 1.9 times higher throughput than serving the target model alone. [5][7][4][5][7][14]
Hardware sets the real floor. The NVFP4 checkpoint targets NVIDIA Blackwell graphics processors. Two-bit quantizations run on a single Apple Silicon Ultra machine with 512 gigabytes of unified memory according to the llama.cpp team. Other configurations need distributed inference across several cards. A trillion-parameter model remains a server-room system for most teams. [5][7][4][5][7][14]
Benchmark bills exclude the rest of the cost. Artificial Analysis notes its task costs do not measure self-hosting or customization or operational oversight or human review. Enterprises still need to compare the cost of work completed to an acceptable standard inside their own environments. The download is free. Running it well is not. [4][14][4][5][7][14]
What to watch next
A full technical report is still coming. Mistral says it will publish deeper detail on architecture and benchmarks and post-training method. Until that report lands, several figures rest on Mistral's own testing and early partner access. Independent researchers can now run the public weights themselves. Their scores will test the announcement within weeks. [1][8][1][4][8][9][15]
The licence clause will face its first real test in practice. European hosters must decide whether to pay for a commercial licence or route around it to rival open models. The boundary between embedding a model in a product and operating a model service remains undefined in the announcement. How Mistral enforces that line will shape who builds on Large 4. [4][9][1][4][8][9][15]
The release date itself carries a lesson. Reuters reported 27 October for the weights. The Hugging Face card counted down to 31 October. Mistral's own announcement promised only the end of the month. The weights landed on 29 October with the licence file. Teams that plan around a document rather than a rumoured date will avoid the same wait. [15][1][1][4][8][9][15]
Questions, answered.
What does open weight mean here?
Open weight means the trained parameters are published for download. Anyone can inspect the model and fine-tune it and run it on their own infrastructure. It does not mean the training data is published. It does not mean every component needed to reproduce training from scratch is included. The model remains available through Mistral's API for teams that prefer a hosted service.[1][3]
Is Large 4 really the best open model for coding?
Mistral says so on its own benchmarks. Independent testing ranks it second on the Coding Agent Index, one point behind GLM-5.3. The one-point gap alone does not establish statistical significance. On broader capability the model scores 26 on Intelligence Index v5. That leads open-weight models developed outside China. Chinese open models score higher at 32. The preview scored 38 on the older v4.3 scale. The two scales are not comparable. The like-for-like v5 comparison is 25 to 26.[1][4]
Can I run it on my own computer?
Most readers cannot run the full model on a laptop. The NVFP4 checkpoint targets datacenter-class NVIDIA Blackwell hardware. Two-bit quantizations run on a single Apple Silicon Ultra machine with 512 gigabytes of unified memory. Other setups need distributed inference across several cards. The API offers a cheaper route for individuals. Input costs $0.70 per million tokens and output costs $2.10.[5][7]
What does the licence forbid?
The licence permits use and modification and fine-tuning and commercialization. It forbids commercial use without a separate agreement for businesses that operate a model-as-a-service offering. Internal use is exempt. Products that embed model capability inside specific features are excluded from the definition. Services that merely relay requests to models hosted elsewhere are excluded too. Large 3 shipped under Apache 2.0. Large 4 does not carry that licence.[4][9][12]
Why give away a cyber model?
Mistral argues defenders need the same tools as attackers under their own control. A provider's refusal filters can block legitimate vulnerability research. A hosted service can vanish mid-incident. Pierre Stock told TechCrunch that open weights give organisations access that cannot be cut off. Sophos threat intelligence director Rafe Pilling calls the release clearly dual-use technology. He says every advanced model needs careful supervision no matter its country of origin.[1][3][14]
What is still unknown?
The full technical report has not arrived. Independent scores on the public weights are still early. The licence boundary between embedding a model and operating a model service is undefined. Mistral has not said how it will enforce the model-as-a-service condition. The pen-testing variant remains restricted to vetted partners. Its wider availability is unconfirmed.[1][8][12]
Read the sources.
- Mistral Large 4 (Le Chonk) is now open weightMistral · Accessed 2026-11-05
- Introducing Mistral Large 4Mistral · Accessed 2026-11-05
- Mistral releases weights for its trillion-parameter AI model, targeting cybersecurityTechCrunch · Accessed 2026-11-05
- Mistral Large 4 'Le Chonk' arrives as the top open model outside China with improved coding performance and 63% lower cost-per-taskVentureBeat · Accessed 2026-11-05
- Welcome Le Chonk: Mistral Large 4 on Hugging FaceHugging Face · Accessed 2026-11-05
- Mistral opens Large 4 weights for coding and securityTesting Catalog · Accessed 2026-11-05
- Mistral's 1-Trillion-Parameter Le Chonk Checkpoint Lands on Hugging FaceUnite.AI · Accessed 2026-11-05
- Mistral Large 4 open weights: specs, license, cyber claimsSkimwire · Accessed 2026-11-05
- Mistral Opens 1T 'Le Chonk' Large 4 Under Modified MIT LicenseAI Weekly · Accessed 2026-11-05
- Mistral opens the weights of its 1-trillion-parameter Mistral Large 4Pivot News · Accessed 2026-11-05
- Mistral Large 4 goes open weight — DeepSeek still beats it on cyberSiliconNoon · Accessed 2026-11-05
- Mistral Releases Mistral Large 4 for Self-Hosted Cyber DefenseWinbuzzer · Accessed 2026-11-05
- Mistral Large 4 Is Now Free to Download: a 1-Trillion-Parameter Model You Can Run YourselfTechFeatured · Accessed 2026-11-05
- Mistral Large 4 Weights Released: How Le Chonk Scores Against RivalsImplicator · Accessed 2026-11-05
- The Reported Date Arrived and the Weights Did NotForkast · Accessed 2026-11-05
Contributor articles and source links do not guarantee error-free reporting. See our editorial policy.
