{"slug": "mistral-large-4-1t-parameters-61-7-deepswe-2026", "title": "Mistral Large 4: 1T Parameters, 61.7% DeepSWE [2026]", "summary": "Mistral AI released the open weights of Mistral Large 4 on Hugging Face on October 29, 2026, three weeks after opening a public preview through its API on October 6, according to Fortune. The Paris-based lab claims the 1-trillion-parameter, multimodal Mixture-of-Experts model with a 1-million-token context window is the strongest open-weight model for agentic coding and cyber defense, citing scores of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4, 51% on the Artificial Analysis Cyber Index, and resistance to 93.3% of attacks on Lakera's public B3 AI Security Benchmark; independent testing has not yet confirmed the claim. The model ships under the Mistral Ecosystem Open-Weight license, described as a modified MIT license, with Orange Cyberdefense named as design partner and launch pricing at 50% off for two weeks.", "body_md": "Mistral AI says its new flagship, Mistral Large 4, is the strongest open-weight model available for agentic coding and cyber defense. [Fortune reported on October 29, 2026](https://fortune.com/2026/10/29/mistral-releases-new-model-large-4-and-claims-top-open-model-for-agentic-coding-and-cyber-defense/) that the Paris-based lab has put the 1 trillion-parameter model on Hugging Face, three weeks after it opened a public preview through its API. The claim is Mistral’s own, and independent testing has not yet settled it.\n\nThe numbers Mistral published on October 6 are specific enough to test. The model scored 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4. On security, it hit 51% on the Artificial Analysis Cyber Index and resisted 93.3% of attacks on Lakera’s public B3 AI Security Benchmark. Mistral also reports a 96.6% average refusal rate on malicious cyber requests.\n\nThis analysis walks through what is confirmed, what is claimed and what nobody has verified yet. It also places the release inside the wider contest between Western and Chinese open-weight labs, a contest we have tracked in our [open-weight license comparison of six model families](https://shattered.io/open-weight-models-license-comparison-6-families-16gb-2026/). Every figure below comes from Mistral’s announcement or from named outlets covering it. Where the record is thin, the article says so.\n\nDon't miss new tech stories on Google\n\nAdd Shattered once in the Google app and our stories show up in your news suggestions.\n\n[Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F)\n\n## What Mistral Large 4 Is, in Numbers\n\nMistral describes Large 4 as an open-weight, general-purpose, multimodal model built on a granular Mixture-of-Experts architecture. The headline size is 1 trillion parameters, and the context window stretches to 1 million tokens. A Mixture-of-Experts design routes each token through a subset of the network instead of the whole thing, which keeps inference cost well below what a dense model of the same size would demand.\n\nThe word “multimodal” matters for buyers. Large 4 is not a text-only coder. Mistral lists the target workloads as agentic tasks, mathematics, reasoning, coding, cybersecurity, manufacturing, spatial reasoning, financial tasks and multimodal usage. That spread tells you where the company wants revenue. It wants engineering teams, security operations centers, factory floors and finance desks, all running a model they can host themselves.\n\nThe model carries the Mistral Ecosystem Open-Weight license, which Mistral calls a modified MIT license. That phrase signals permissive terms, but “modified” is the operative word. Legal teams will want to read the actual text before they ship a product on top of it. We return to that question in the licensing section below.\n\n| Attribute | Mistral Large 4 (as announced) | \n|---|---|\n| Developer | Mistral AI | \n| Architecture | Granular Mixture-of-Experts, multimodal | \n| Total parameters | 1 trillion | \n| Context window | 1 million tokens | \n| Preview announced | October 6, 2026 | \n| License | Mistral Ecosystem Open-Weight (described as modified MIT) | \n| Named design partner | Orange Cyberdefense | \n| Launch pricing | 50% off for 2 weeks (base price not confirmed) | \n\n## Timeline: Preview on October 6, Weights on October 29\n\nMistral opened the public preview on October 6, 2026 and pointed developers to its Studio API. The company also promised that the weights would arrive by the end of the month. It kept that promise. Coverage from Fortune and The Next Web on October 29 says the weights are now downloadable, and The Next Web describes the released version as noticeably sharper than the preview.\n\nThat gap between preview and release is the part worth studying. Mistral did not merely flip a switch on files it had finished weeks earlier. It kept training. Its science team says the open-weight build continued reinforcement learning after the preview, and the company claims the gains showed no signs of saturating. The scores in the October 6 announcement therefore describe the preview, not necessarily the model now sitting on Hugging Face.\n\nThis distinction trips up casual readers. When a lab says “here are our benchmarks” on day one and ships a better checkpoint on day twenty-three, the earlier table is a floor. Independent evaluators need to rerun the suite against the final weights before anyone declares a leaderboard winner.\n\n| Date | Event | Source type | \n|---|---|---|\n| September 29, 2026 | OpenAI introduces GPT-6.1 Sol for coding, computer use and professional work | OpenAI release page | \n| October 6, 2026 | Mistral announces Large 4 in public preview, with benchmark scores | Mistral announcement | \n| Launch period (exact dates not confirmed) | 50% off for 2 weeks on launch pricing | Mistral changelog (reported) | \n| Late October 2026 | Mistral promises open weights by month end | Mistral announcement | \n| October 29, 2026 | Weights go live on Hugging Face, per Fortune | Fortune, The Next Web | \n\n## Coding Benchmarks: 61.7%, 59.4% and 28.3% Explained\n\nThree coding scores anchor Mistral’s pitch. DeepSWE v1.1 came in at 61.7%. SWE-Atlas-QnA landed at 59.4%. Terminal-Bench 4 produced 28.3%. The first two sit in the high fifties and low sixties, which Mistral presents as frontier-class for an open model. The third is much lower, and it deserves a plain reading.\n\nTerminal-Bench 4 targets terminal-based coding and computer-use tasks, the kind of long, multi-step work an autonomous agent performs when it has to run commands, read output and recover from errors. A 28.3% score means the preview model finished roughly one in four such tasks. That is a hard benchmark, and low absolute numbers are normal there. Still, it shows that agentic reliability, not single-file code generation, is where the gap between models shows up.\n\nMistral has not published a head-to-head table against named rivals in the material we reviewed, so we cannot say how 61.7% on DeepSWE v1.1 compares with a competitor’s score on the same test. We will not guess. What we can say is that the company chose to lead with coding evaluations built around software-engineering agents, which fits the product positioning. Developers who compare options can start with our [coding-assistant test harness](https://shattered.io/best-ai-for-coding-free-test-harness-12-steps-2026/) and feed the same prompts to each candidate.\n\n### How to read a lab’s own benchmark table\n\nSelf-reported scores carry three standard caveats. The lab picks which benchmarks to publish. It controls the scaffolding, the prompt format and the number of attempts. And it can tune training toward the test set, deliberately or not. None of that makes Mistral’s numbers wrong. It means they are a starting hypothesis, and the community’s reruns will carry more weight in a month than the press release does today.\n\n| Benchmark | Domain | Mistral Large 4 score | Reported in | \n|---|---|---|---|\n| DeepSWE v1.1 | Software engineering | 61.7% | Oct. 6 announcement | \n| SWE-Atlas-QnA | Code question answering | 59.4% | Oct. 6 announcement | \n| Terminal-Bench 4 | Terminal and agentic tasks | 28.3% | Oct. 6 announcement | \n| Artificial Analysis Cyber Index | Cybersecurity | 51% | Mistral, as reported | \n| Lakera B3 AI Security Benchmark | Prompt-injection resistance | 93.3% | Mistral, as reported | \n| Malicious cyber requests | Refusal rate (average) | 96.6% | Mistral, as reported | \n\n## Cyber Defense Claims: 51% Index Score and 96.6% Refusals\n\nThe cyber-defense angle is what separates this launch from a routine coding-model update. Mistral reports 51% on the Artificial Analysis Cyber Index and says the model refuses malicious cyber requests 96.6% of the time on average. Fortune frames the model as the best open-weight option on cyber in particular, and on coding, which is Mistral’s own description of the release.\n\nThose two numbers pull in opposite directions, and the tension is the story. A defender wants a model that reads malware, triages alerts and writes detection rules. An attacker wants the same skills pointed the other way. A high refusal rate suggests Mistral tried to shape behavior on the hosted API. But refusal training lives in the weights, and open weights can be fine-tuned by anyone with enough compute. A refusal rate measured on the stock model says little about a modified copy.\n\nSecurity teams already juggle a heavy flow of exploited-flaw advisories, as our look at [InfraTrust’s October pulse of 16 KEV advisories](https://shattered.io/infratrust-october-2026-pulse-16-kev-advisories-edge-devices-2026/) shows. A self-hosted model that triages that flow without sending telemetry to a third party is an attractive pitch, especially for regulated sectors.\n\n### Prompt-injection resistance: 93.3% on Lakera’s B3\n\nAgents read untrusted text all day. They scrape web pages, parse emails and open tickets, and any of those can carry hidden instructions. Mistral says Large 4 resisted 93.3% of attacks on Lakera’s public B3 AI Security Benchmark. That is a respectable score, and it also means roughly one attack in fifteen got through. For an agent with shell access, one in fifteen is not a comfortable margin. Teams should still wrap the model in least-privilege sandboxes and human approval gates, whatever the benchmark says.\n\n### Orange Cyberdefense as design partner\n\nMistral named Orange Cyberdefense as an enterprise design partner. The detail matters because it ties the model to operational security work instead of lab-only tests. The available coverage does not describe the scope of the collaboration, the size of any deployment or any commercial terms, so readers should treat the partnership as a signal of intent rather than proof of production use.\n\n## Open Weights and the License Question\n\nOpen-weight does not mean open-source, and the industry keeps blurring the two. With Large 4, Mistral publishes the trained weights so anyone can download, host and adapt the model. It does not, based on the coverage we reviewed, publish the training data or the full training pipeline. That is the standard definition of open-weight, and it is the same bargain most major labs now offer.\n\nThe Mistral Ecosystem Open-Weight license is where the practical details sit. Mistral calls it a modified MIT license. A plain MIT license grants broad rights with almost no conditions. A modified version can add carve-outs for hosting competitors, revenue thresholds, acceptable-use rules or attribution terms. We have not reproduced those terms here because the fact sheet behind this report does not list them. Anyone planning a commercial product should read the license file on the model repository directly.\n\nOur [comparison of DeepSeek, Qwen and Mistral open-weight models](https://shattered.io/deepseek-vs-qwen-vs-mistral-open-weight-models-2026/) found that license wording often matters more to procurement teams than a two-point benchmark gap. A model that legal approves in a week beats a slightly stronger one that stalls in review for a quarter.\n\n## What Mistral’s Science Chief Said\n\nPierre Stock, Vice President of Science at Mistral, has been the company’s main voice on the release. His comments to the press touch three themes: why open weights matter, what happens once weights are public, and how the post-preview training went. The quotes below are verbatim from the cited reports.\n\nOn the philosophy behind shipping weights at all, Stock told [Reuters](https://www.reuters.com/legal/litigation/mistral-ai-publicly-releases-large-4-targets-cybersecurity-users-2026-10-29/): “I’m not comfortable living in a world in which open weights do not exist.” It is a values statement, and it explains why Mistral keeps releasing weights even as rivals lock their best models behind APIs.\n\nOn the risk side, Stock acknowledged the tradeoff that comes with a cyber-capable model. In the same Reuters report he said: “Once the model is out, it’s impossible to contain access”. That candor is useful. It concedes that the refusal rate and safety tuning cannot recall a download, and it puts the burden of responsible deployment on the people who run the weights.\n\nOn training progress, [Euronews](https://www.euronews.com/2026/10/29/mistral-releases-latest-ai-model-for-download-giving-companies-more-control-over-their-cy) quotes Stock as saying: “Building on the ML4 Preview, the open weight version continues reinforcement learning training and delivers substantial gains with no signs of saturation”. He added that “ML4 OSS showcases our science team’s ability to innovate at speed.” The first line is the more important one for buyers, because it implies the model will keep improving and that the October 6 scores understate the shipped checkpoint.\n\nTaken together, the remarks sketch a company that sees open weights as both a business strategy and a principle. They also show Mistral choosing not to hide the dual-use problem, which is the right call for a launch aimed at security buyers.\n\n## Competitive Picture: Closed Frontier vs Open Weight\n\nMistral is not launching into an empty field. OpenAI introduced GPT-6 Sol and Luna, then followed with GPT-6.1 Sol on September 29, 2026. Its release page pitches GPT-6.1 Sol as near its top-tier Astra model for coding, computer use and professional work, at one-fifth of Astra’s standard API prices. That is a direct shot at the cost-conscious developers Mistral wants to win. Our report on [GPT-6.1 Terra in ChatGPT Work and Codex](https://shattered.io/gpt-6-1-terra-chatgpt-work-codex-api-pricing-2026/) covers the pricing side of that fight.\n\nRelease trackers also list Kimi K3.1 Pro, Gemini 4 and Muse Wonder among the frontier releases of 2026. We do not have verified scores for those models on the same benchmarks Mistral used, so a clean scoreboard is impossible today. The Yahoo Tech report on the October 6 preview says Mistral claims Large 4 outperforms every rival built outside China, a carefully worded boundary that concedes the top Chinese open models may still lead.\n\n| Model or group | Lab | Access type | What is verified here | \n|---|---|---|---|\n| Mistral Large 4 | Mistral AI | Open weights plus API | 1T parameters, 1M-token context, scores listed above | \n| GPT-6.1 Sol | OpenAI | API product | Released Sept. 29, 2026, priced at one-fifth of Astra’s standard rates | \n| GPT-6 Sol and Luna | OpenAI | API product | Introduced before GPT-6.1 Sol, frontier-tier positioning | \n| Kimi K3.1 Pro | Moonshot | Not confirmed here | Tracked as a 2026 frontier release, no shared benchmark data | \n| Gemini 4 |  | Not confirmed here | Tracked as a 2026 frontier release, no shared benchmark data | \n| Muse Wonder | Not confirmed here | Not confirmed here | Tracked as a 2026 frontier release, no shared benchmark data | \n\nThe empty cells are honest. Fortune and the trackers confirm the rivals exist and shipped this year. They do not supply the head-to-head numbers a proper comparison needs. For readers choosing among open families today, our roundup of the [best open-source LLMs of October 2026](https://shattered.io/best-open-source-llm-october-2026-kimi-glm-deepseek/) offers a wider field.\n\n## The China Factor and Western Open-Weight Anxiety\n\nFortune’s framing is that Mistral has chased its Chinese rivals to the frontier of open-weight AI. That sentence captures the geopolitics baked into this release. Over the past year, Chinese labs set much of the pace in open weights, and Western enterprises that wanted to self-host a strong model often had to pick a Chinese one. Governments noticed. Our coverage of the [DHS DeepSeek order and China’s open-weight strategy](https://shattered.io/china-open-weight-ai-strategy-dhs-deepseek-order-oct-20-2026/) shows how quickly procurement rules can turn on a model’s country of origin.\n\nA trillion-parameter model trained in European data centres offers an alternative that clears many of those rules by default. Mistral’s earlier public preview coverage emphasized that training happened on its own European infrastructure, a point that resonates with buyers worried about data residency and regulatory exposure. For a bank in Frankfurt or a defense contractor in Paris, “Made in Europe” can be a deciding feature before any benchmark enters the conversation.\n\nThe security argument runs both ways. Some analysts see open Chinese models as a risk, others as a defensive resource, a debate we unpacked in [China’s open AI: three labs and one security debate](https://shattered.io/china-open-weight-ai-weapon-or-armour-cybersecurity-debate-2026/). Large 4 adds a Western entrant to that argument and gives policymakers a new data point.\n\n## Market Impact for Enterprises and Developers\n\nThe practical impact lands in three places. First, self-hosting economics. A Mixture-of-Experts model with 1 trillion parameters still needs serious GPU memory, so the weights are not a laptop download. Large enterprises and cloud providers will move first, and smaller teams will most likely consume Large 4 through hosted endpoints. That keeps Mistral’s API business alive even as the weights go public.\n\nSecond, pricing pressure. Mistral used a 50% launch discount for two weeks during the preview. The underlying list price is not confirmed in the changelog we reviewed, so we cannot say how Large 4 compares per token with GPT-6.1 Sol. The discount itself is the signal. Mistral is willing to buy adoption, and rivals with closed models must decide whether to answer with price cuts of their own.\n\nThird, procurement leverage. Once a competitive open-weight model exists, every buyer negotiating with a closed-model vendor gains a credible alternative. Even teams that never deploy Large 4 benefit from the option.\n\n## Historical Context: Mistral’s Open-Weight Playbook\n\nMistral built its reputation by shipping capable open-weight models while larger American labs guarded theirs. The strategy gave a young European company global developer mindshare at a fraction of the marketing budget. Large 4 is the largest expression of that playbook so far, scaling the open release up to the 1 trillion-parameter tier that once belonged only to closed systems.\n\nThe wider market has shifted underneath it. One tracker, The Durability Curve, counts 53 frontier releases from ten labs in the year to October 2026, up from 23 the year before. That pace compresses every launch window. A model that leads on October 29 may trail by Thanksgiving, which is why Mistral’s decision to keep reinforcement learning running after the preview looks less like polish and more like survival.\n\nThe same acceleration explains why labs now fight over safety framing. Anthropic urged courts to block risky models in an October letter, covered in our [report on Anthropic’s call for courts to block risky AI](https://shattered.io/anthropic-newsom-letter-courts-block-risky-ai-models-oct-15-2026/). Mistral’s open release of a cyber-capable trillion-parameter model sits at the opposite end of that policy spectrum, and regulators in Brussels and Washington will likely read the two positions side by side.\n\n## What Remains Unconfirmed\n\nSeveral points stay unproven, and we flag them so readers do not mistake marketing for settled fact. The claim that Large 4 is the best open model overall is Mistral’s performance claim, recorded by Fortune and others, not an independently established ranking. Fortune’s own wording ties the best-in-class label to agentic coding and cyber defense specifically.\n\nThe launch price is also unconfirmed. The official changelog says 50% off for two weeks, but it does not state the base price in the material we reviewed. Third-party benchmark reproductions of the coding and cyber scores had not appeared at the time of writing. And the scores Mistral published on October 6 describe the preview, so the final weights could land higher or, less likely, differently on some tests.\n\nFinally, the safety numbers deserve scrutiny. A 96.6% refusal rate and a 93.3% injection-resistance score are measurements of one model state under one evaluation setup. They say nothing about fine-tuned derivatives, which the open license allows. Security researchers will probably test how quickly refusals can be stripped, and that result will shape how enterprises and regulators judge the release.\n\n## Five Predictions for Mistral Large 4\n\nThese are our forecasts, built from the facts above. They are analysis, not reported events.\n\n- **Independent reruns will land within weeks.** Open weights invite community evaluation, so expect third-party scores on DeepSWE and Terminal-Bench to appear before mid-November, probably with modest gaps from Mistral’s numbers.\n- **Terminal-Bench 4 becomes the scoreboard that matters.** The 28.3% preview result is the weakest of Mistral’s coding scores, and the next training round will target it.\n- **A fine-tuned, safety-stripped variant will surface.** Open cyber-capable weights almost always spawn uncensored derivatives, which will force a public debate on the 96.6% refusal figure.\n- **Rivals will answer on price.** With Mistral discounting and giving away weights, closed-model vendors will sharpen their API pricing for coding workloads.\n- **European buyers will pilot it first.** Data-residency rules and the Orange Cyberdefense tie-in make regulated European sectors the likeliest early adopters.\n\n## How Teams Should Evaluate It This Week\n\nA sensible evaluation starts with your own tasks. Pull 30 to 50 real tickets, failing tests or security alerts from your backlog, and run them through Large 4, a closed frontier model and whatever you use today. Score the outputs blind. Public benchmarks tell you about the model in general, while your own repository tells you about the model in your stack.\n\nNext, test the agent loop and not just single answers. Give the model a sandbox with a terminal and a time budget, then measure how often it completes a task unaided. Track failure modes: wrong commands, looping, ignoring test output. Terminal-Bench 4’s modest 28.3% suggests this is where differences will show.\n\nFinally, red-team the safety claims on your own terms. Plant hostile instructions in documents the agent reads and see whether it follows them. Review the license with counsel. And price the infrastructure honestly, because a 1 trillion-parameter model changes your GPU budget even if the weights cost nothing to download.\n\nFor wider context on model categories and where this release fits, see our [AI and machine learning coverage hub](https://shattered.io/ai-machine-learning/).\n\n## FAQ: Mistral Large 4\n\n### When did Mistral announce Mistral Large 4?\n\nMistral announced a public preview on October 6, 2026. Fortune and The Next Web reported on October 29 that the open weights are now available on Hugging Face.\n\n### How large is the model and what is its context window?\n\nMistral lists 1 trillion parameters in a granular Mixture-of-Experts design and a context window of 1 million tokens. It is multimodal and general-purpose.\n\n### What are the coding benchmark scores?\n\nThe October 6 announcement reported 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4. These describe the preview, and the final weights received additional training.\n\n### Is it really the best open model?\n\nThat is Mistral’s claim, focused on agentic coding and cyber defense. It is not an independently established ranking, and Mistral’s wording about outperforming rivals carves out models built in China.\n\n### What cybersecurity results did Mistral report?\n\nMistral reports 51% on the Artificial Analysis Cyber Index, 93.3% attack resistance on Lakera’s B3 benchmark and a 96.6% average refusal rate on malicious cyber requests. Orange Cyberdefense is a named design partner.\n\n### What license covers the weights?\n\nThe Mistral Ecosystem Open-Weight license, which Mistral describes as a modified MIT license. Read the license file itself before building a commercial product.\n\n### How much does it cost?\n\nThe official changelog confirms only a launch discount of 50% off for two weeks. A base price is not confirmed in the material we reviewed, and the weights themselves can be downloaded without a license fee, though hosting costs apply.", "url": "https://wpnews.pro/news/mistral-large-4-1t-parameters-61-7-deepswe-2026", "canonical_source": "https://shattered.io/mistral-large-4-open-weight-1t-parameters-deepswe-61-7-2026/", "published_at": "2026-10-29 16:27:30+00:00", "updated_at": "2026-10-29 16:50:17.372354+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-safety", "ai-research", "ai-products"], "entities": ["Mistral AI", "Mistral Large 4", "Hugging Face", "Fortune", "Orange Cyberdefense", "Lakera", "Artificial Analysis", "The Next Web"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/mistral-large-4-1t-parameters-61-7-deepswe-2026", "markdown": "https://wpnews.pro/news/mistral-large-4-1t-parameters-61-7-deepswe-2026.md", "text": "https://wpnews.pro/news/mistral-large-4-1t-parameters-61-7-deepswe-2026.txt", "jsonld": "https://wpnews.pro/news/mistral-large-4-1t-parameters-61-7-deepswe-2026.jsonld"}}