--- title: Mistral releases open weights for 1 trillion-parameter... description: Mistral released open weights for Mistral Large 4, a 1 trillion-parameter mixture-of-experts multimodal model, on Hugging Face. It activates about 52 billion... canonical: https://daily.dev/posts/mistral-releases-open-weights-for-1-trillion-parameter-large-4-with-day-0-vllm-support-hrgscxwdb twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Mistral releases open weights for 1 trillion-parameter Large 4, with day-0 vLLM support | daily.dev og:description: Mistral released open weights for Mistral Large 4, a 1 trillion-parameter mixture-of-experts multimodal model, on Hugging Face. It activates about 52 billion... og:url: https://daily.dev/posts/mistral-releases-open-weights-for-1-trillion-parameter-large-4-with-day-0-vllm-support-hrgscxwdb og:image: https://api.daily.dev/og/posts/HrGSCXWDb.png og:image:alt: Mistral releases open weights for 1 trillion-parameter Large 4, with day-0 vLLM support og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Mistral releases open weights for 1 trillion-parameter Large 4, with day-0 vLLM support **[Collections](https://daily.dev/sources/collections)** · 4 min read · 9 upvotes · 0 comments ## Summary Mistral released open weights for Mistral Large 4, a 1 trillion-parameter mixture-of-experts multimodal model, on Hugging Face. It activates about 52 billion parameters per task, per Mistral, or 49 billion per token per vLLM's figures. On Terminal-Bench 4.0, extra reinforcement learning pushed its agentic coding score from 28% to 41%, edging GLM-5.3's 40% while using 47% fewer tokens per task. It scored 79% on CyberGym-E2E, matching MiMo-V2.6-Pro and beating DeepSeek V4.1 Flash's 76%. Mistral trained it on 3,800 Nvidia Grace Blackwell GPUs and has raised roughly $6 billion, CEO Arthur Mensch said, versus hundreds of billions at other labs. vLLM supports it day-0, with sparse MLA attention, cross-layer KV sharing, a DSpark speculative decoder giving about 2x throughput on TP8, and day-0 support on AMD MI350 and TPU v7x too. ## Content Mistral has released the weights of Mistral Large 4, nicknamed Le Chonk, on Hugging Face. It is a 1-trillion-parameter multimodal mixture-of-experts model, previewed earlier this month. The license is a modified MIT license, not the Apache 2.0 that covered Large 3. ## License and pricing The Mistral Ecosystem Open-Weight license lets anyone use, modify and commercialize the model. Model-as-a-Service businesses, though, must get a separate commercial license. That is tighter than Large 3's terms. API pricing is now $0.70 per million input tokens, $0.07 for cached input and $2.10 for output. By one comparison that is about half the list prices of $1.36, $0.14 and $4.18 that applied by 22 October. Vercel's AI Gateway has switched to the updated weights and lower prices. Existing users of the `mistral/mistral-large-4` model ID get both automatically. Agent harnesses such as Claude Code, Codex and Cursor can connect by installing the latest Vercel CLI and running `npx vercel ai-gateway setup`. The gateway charges no markup on inference, including BYOK requests. ## Benchmarks Mistral says three more weeks of reinforcement learning since the preview lifted its Terminal-Bench 4.0 agentic coding score from 28% to 41%. That is one point above GLM-5.3 at 40%. It also scores 19.7% on Cognition's FrontierCode 2.0, which Mistral says leads open-weight models. The lead is thin. On Artificial Analysis's Coding Agent Index, Large 4 scores 48%, effectively a tie with GLM-5.3. Its edge is efficiency: it uses 47% fewer tokens per task on average. It activates 52 billion parameters per task. On CyberGym-E2E, which tests finding and patching memory-safety bugs, it scored 79%. That matches MiMo-V2.6-Pro and beats DeepSeek V4.1 Flash at 76%. Mistral also reports 51% on the Artificial Analysis Cyber Index. Training ran on 3,800 Nvidia Grace Blackwell GPUs in Europe. CEO Arthur Mensch said Mistral has raised roughly $6 billion in three years, against hundreds of billions at other labs. VP of Science Pierre Stock said the RL training used just 10 megawatts of power, and that smaller models, including a laptop-sized one, are coming. ## The cyber question Mistral open-sourced the model while warning it could help attackers carry out real cyberattacks. The company says it refused 96.6% of malicious prompts in the cyber portions of JailbreakBench, StrongREJECT and AgentHarm, the highest rate it has measured among open models. There is a catch. Safeguard updates Mistral issues later will apply only to its managed API and future releases, not to checkpoints people have already downloaded. A UK AI Security Institute report found open-weight models trail closed frontier models on cyber tasks by four to seven months. A cybersecurity-focused version is being tested by a small group of partners, restricted to authorized systems. Orange Cyberdefense is the first European managed security services provider to join the design partner programme. It plans to use Large 4 for technical audits, penetration testing, breach investigation and incident response, running it in Mistral's European data centers so clients keep control of their data. ## Day-0 tooling The ecosystem moved quickly. - **vLLM** supports the model with FP8 and NVFP4 checkpoints and a built-in DSpark speculative decoding drafter, which gives about 2x output throughput on TP8. Decode context parallelism gives roughly 7.9x more KV capacity on TP8. NVFP4 matches FP8 quality within 2.1 points and decodes up to 1.22x faster. A HiSparse host-memory KV offload holds up to 5.7x more KV tokens than GPU memory alone. It also runs on AMD Instinct MI350 GPUs and TPU v7x. - **RL with SkyRL and vLLM:** on 8 Nvidia GB200 nodes, LoRA RL on the NVFP4 checkpoint lifted AIME 2024 avg@8 from about 64% to about 75% in 75 steps. The recipe uses rank-32 LoRA, NVFP4 quantization-aware training, FP8 KV-cache simulation, routing replay and token-level truncated importance sampling. The FSDP2 run kept a mean absolute trainer-vLLM logprob gap of about 0.02 to 0.026, and Megatron reached similar accuracy. - **Axolotl** supports LoRA fine-tuning. Its recommended recipe keeps experts in NVFP4 with merge-aware LoRA, and fits on one 8xB200 node (8xB300 is ideal). - **Unsloth** made quants. Its 1-bit version keeps 91% of top-1 accuracy across coding, agentic tasks and math while being 80% smaller, 223GB against 1.1TB for FP8. They work in Unsloth Desktop. - **Transformers and llama.cpp** have day-0 support too. The pairing of a mostly tied benchmark lead with a clearer token-efficiency advantage and a halved API price makes this a practical release more than a flashy one. The license change and the cyber caveat are what I'd read closely before building on it. ## Questions this post answers ### How many parameters does Mistral Large 4 activate per token? Mistral Large 4 is a 1 trillion-parameter mixture-of-experts model that activates only a fraction of its parameters per task. Mistral states 52 billion active parameters per task, while vLLM's day-0 technical write-up reports 49 billion active parameters per token, reflecting slightly different measurement approaches. _Engineers sizing inference hardware for huge MoE models can track releases like this on daily.dev._ ### How does Mistral Large 4 compare to GLM-5.3 on coding benchmarks? Mistral Large 4 scores 41% on Terminal-Bench 4.0 after three extra weeks of reinforcement learning, versus 40% for GLM-5.3. Mistral says Large 4 also uses 47% fewer tokens per task on average than GLM-5.3, meaning comparable coding performance at lower inference cost. _Teams picking between coding models can compare benchmark trade-offs like this via daily.dev._ ### What speedups does vLLM provide for serving Mistral Large 4? vLLM supports Mistral Large 4 from day zero using sparse MLA attention, cross-layer KV sharing, and a built-in DSpark speculative decoding drafter that roughly doubles output throughput on TP8. Additional gains include about 7.9x more KV capacity via decode context parallelism, NVFP4 quantization decoding up to 1.22x faster than FP8 with only a 2.1-point quality gap, and a HiSparse host-memory KV offload holding up to 5.7x more KV tokens than GPU memory alone. _Developers tuning LLM serving throughput can follow inference optimizations like these on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 114 comments across hackernews (as of 2026-10-30). **TL;DR:** Most of the debate centers on whether 'open weight' is a misleading term given Mistral's commercial-use license restriction, while a separate thread of discussion questions whether the model is actually competitive against Chinese and US rivals despite its aggressive pricing. **Sentiment:** 20% positive · 45% mixed · 35% skeptical **The case for** - The price cut from preview pricing (down to $0.70/$2.10 per million tokens) makes it notably more interesting than before. - Some find it a capable, usable option for everyday coding tasks, especially for EU-based users wanting EU provenance. - Mistral adding third-party models like GLM and DeepSeek to its API has made it a solid overall provider for some business users. **The pushback** - The modified MIT license requires a separate commercial license for 'Model as a Service' businesses, which many argue makes 'open weight' a misleading label versus true open source. - Several benchmarks (e.g. Artificial Analysis Intelligence Index) suggest the model isn't cost/performance competitive against rivals. - The 96.6% refusal rate on malicious requests is seen as a weak or easily gamed metric with no corresponding data on false-positive refusals of legitimate requests. - Personal benchmarking and anecdotal testing suggest it lags behind top Chinese and US open-weight models. **By community** - hackernews (heated): A long, contentious debate over whether the license truly qualifies as 'open weight' dominates, alongside a separate skeptical thread about the model's real-world competitiveness. **Hottest debate:** Whether Mistral's license (which restricts commercial 'Model as a Service' use) can honestly be called 'open weight' or should instead be termed 'weights available'. **Open questions** - What is an authoritative definition of 'open weight' when no single body (OSI or otherwise) is universally recognized as the arbiter? - How does Mistral Large 4 actually perform on cost-adjusted intelligence benchmarks compared to competitors like GLM 5.3 and DeepSeek? - Why do malicious-request refusal rate statistics not come paired with false-positive refusal rates on legitimate requests? **Highlights** > Just because OpenAI ran roughshod over the broader software community's notion of "open" which once heavily implied free-as-in-software-not-beer, doesn't mean we need to continue this pattern. Many industries had "source available" without the word "open" for decades, and that was a broadly understood term. "Weights available" would be a reasonable analogue. For clarity, the Mistral model license is here: https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/raw... > 2. If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose. The Licensee may contact Mistral, using this form: https://mistral.ai/contact/, to request a commercial license. > "Model as a Service" means giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints), in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data. This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others. "Meaningful control over the inputs" is so ambiguous as to be nearly meaningless. Does a model merely summarizing a user's submitted support tickets, give that user meaningful control over the inputs? Don't get me wrong - I appreciate that Mistral needs to walk a fine line for its own sustainability, and that sustainability is a good thing! And I may just be an old person ranting at a cloud. But "open" used to mean something, back in my day! > — [btown on hackernews](https://news.ycombinator.com/item?id=50289907) > Ughh the overly positive and cheerful comments here from what is usually the most critical/pessimistic audience on the internet is a huge tell. If we cut through the political bias and wishful thinking for a second, here’s the truth: The model is not competitive. I would never waste my time using it for coding work, and I’m someone in the EU. Great effort, but keep at it. This just isn’t good enough. I’m sure you guys can make a boatload of money with regulatory capture and compliance games in the EU, but please don't give up trying to produce something thats actually competitive. > — [pembrook on hackernews · 3 comments](https://news.ycombinator.com/item?id=50290705) > "Cyber benchmarks, while achieving an average refusal rate of 96.6% on malicious cyber requests, the highest we measure among open models." I feel like stating the obvious here and hence like I'm missing something. Where is the statistic about not refusing the non malicious requests? I can achieve the same performance with refusing when random(1000) < 967. > — [3form on hackernews · 3 comments](https://news.ycombinator.com/item?id=50289898) > Look at the Intelligence Index vs. Cost per Task. Mistral 4 is not competitive. https://artificialanalysis.ai/#intelligence-efficiency-tabs > — [esafak on hackernews · 1 comments](https://news.ycombinator.com/item?id=50290064) > They slashed pricing, it's now nearly half vs the preview (Input: $0.7/M Tokens, Output: $2.1/M Tokens vs Input: $1.36/M Tokens, Output: $4.18/M Tokens). At the original pricing in my evals, Mistral Large 4 Preview was not competitive. With this price point and some improvements on top, this full release warrants a major reassessment. [0] https://docs.mistral.ai/inference/pricing > — [Topfi on hackernews · 4 comments](https://news.ycombinator.com/item?id=50289701) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50289418) · 382 points · 114 comments ## Similar posts on daily.dev - [Building AIE CODE: The top Coding Agent Conference](https://daily.dev/posts/building-aie-code-the-top-coding-agent-conference-ovcg1zp0l) · Latent Space · 0 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#llm](https://daily.dev/tags/llm), [#vllm](https://daily.dev/tags/vllm), [#mistral-ai](https://daily.dev/tags/mistral-ai) [View this post on daily.dev](https://daily.dev/posts/mistral-releases-open-weights-for-1-trillion-parameter-large-4-with-day-0-vllm-support-hrgscxwdb) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Mistral releases open weights for 1 trillion-parameter Large 4, with day-0 vLLM support","url":"https://daily.dev/posts/mistral-releases-open-weights-for-1-trillion-parameter-large-4-with-day-0-vllm-support-hrgscxwdb","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/mistral-releases-open-weights-for-1-trillion-parameter-large-4-with-day-0-vllm-support-hrgscxwdb"},"datePublished":"2026-10-29T14:19:04.063Z","dateModified":"2026-10-30T13:39:31.404Z","description":"Mistral released open weights for Mistral Large 4, a 1 trillion-parameter mixture-of-experts multimodal model, on Hugging Face. It activates about 52 billion...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/94c2826e28ea2ff7118606ee467d3622?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/94c2826e28ea2ff7118606ee467d3622?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/mistral-releases-open-weights-for-1-trillion-parameter-large-4-with-day-0-vllm-support-hrgscxwdb","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":9},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,llm,vllm,mistral-ai","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Mistral releases open weights for 1 trillion-parameter Large 4, with day-0 vLLM support"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/mistral-releases-open-weights-for-1-trillion-parameter-large-4-with-day-0-vllm-support-hrgscxwdb#faq","mainEntity":[{"@type":"Question","name":"How many parameters does Mistral Large 4 activate per token?","acceptedAnswer":{"@type":"Answer","text":"Mistral Large 4 is a 1 trillion-parameter mixture-of-experts model that activates only a fraction of its parameters per task. Mistral states 52 billion active parameters per task, while vLLM's day-0 technical write-up reports 49 billion active parameters per token, reflecting slightly different measurement approaches. Engineers sizing inference hardware for huge MoE models can track releases like this on daily.dev."}},{"@type":"Question","name":"How does Mistral Large 4 compare to GLM-5.3 on coding benchmarks?","acceptedAnswer":{"@type":"Answer","text":"Mistral Large 4 scores 41% on Terminal-Bench 4.0 after three extra weeks of reinforcement learning, versus 40% for GLM-5.3. Mistral says Large 4 also uses 47% fewer tokens per task on average than GLM-5.3, meaning comparable coding performance at lower inference cost. Teams picking between coding models can compare benchmark trade-offs like this via daily.dev."}},{"@type":"Question","name":"What speedups does vLLM provide for serving Mistral Large 4?","acceptedAnswer":{"@type":"Answer","text":"vLLM supports Mistral Large 4 from day zero using sparse MLA attention, cross-layer KV sharing, and a built-in DSpark speculative decoding drafter that roughly doubles output throughput on TP8. Additional gains include about 7.9x more KV capacity via decode context parallelism, NVFP4 quantization decoding up to 1.22x faster than FP8 with only a 2.1-point quality gap, and a HiSparse host-memory KV offload holding up to 5.7x more KV tokens than GPU memory alone. Developers tuning LLM serving throughput can follow inference optimizations like these on daily.dev."}}]} ```