{"id":10349,"date":"2026-11-17T10:59:21","date_gmt":"2026-11-17T01:59:21","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=10349"},"modified":"2026-11-17T11:08:57","modified_gmt":"2026-11-17T02:08:57","slug":"mistral-large-4-preview-2","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/mistral-large-4-preview-2\/","title":{"rendered":"Mistral AI Announces Mistral Large 4 with 1.05T Parameters"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Published<\/td>\n<td>2026-10-06<\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>European AI startup Mistral AI has launched a public preview of &#8220;Mistral Large 4&#8221; (unofficial name: le Chonk), its largest and most powerful flagship model to date. Developed to push the performance limits of open-weight models, this model is a native multimodal model with a total parameter count of 1.05 trillion and 49 billion active parameters. A preview API is currently available via Mistral Studio, and the model weights are scheduled to be released in late October 2026.<\/p>\n<h2>Specifications<\/h2>\n<p>The specifications of &#8220;Mistral Large 4&#8221; based on published materials are as follows:<\/p>\n<ul>\n<li><strong>Total Parameters<\/strong>: 1.05 trillion (1.05T)<\/li>\n<li><strong>Active Parameters<\/strong>: 49 billion (49B)<\/li>\n<li><strong>Vision Encoder<\/strong>: 1.6 billion (1.6B) parameters<\/li>\n<li><strong>Architecture<\/strong>: Fine-grained Mixture-of-Experts (MoE)<\/li>\n<li><strong>Context Length<\/strong>: 1 million (1M) tokens<\/li>\n<li><strong>Supported Languages<\/strong>: Over 160 languages, including all official languages of the European Union (EU)<\/li>\n<\/ul>\n<h2>Performance<\/h2>\n<p>According to measurement results published by Mistral AI, Mistral Large 4 demonstrates exceptionally high performance in cybersecurity, agent workflows, coding, and multimodal comprehension, surpassing existing powerful open-weight models and some closed models.<\/p>\n<h3>Cybersecurity and Safety<\/h3>\n<p>In the cybersecurity field, the model shows outstanding results. It ranks in the top 5 globally on Artificial Analysis&#8217;s &#8220;Cyber Index&#8221; (an independent evaluation organization), leading open-weight models developed outside of China by a significant margin.<\/p>\n<ul>\n<li><strong>Vulnerability Replication and Patching<\/strong>: It recorded a top score of 82% in tests reproducing and fixing real open-source software vulnerabilities. While competing closed models such as Claude Opus 5.5 and GPT-6 Astra remained at nearly 0% due to safety filter refusals, it demonstrated an overwhelming advantage in practical defensive activities.<\/li>\n<li><strong>Cybench<\/strong>: It solved 93% of the 40-question security competition benchmark, reporting the highest level of scores for an open-weight model.<\/li>\n<li><strong>B3 AI Security Benchmark<\/strong>: In Lakera&#8217;s public benchmark, it blocked 93.3% of attacks, showing the highest resistance among competing models.<\/li>\n<\/ul>\n<h3>Agents and Coding<\/h3>\n<p>Its capabilities as an agent for development and business automation are also highly evaluated.<\/p>\n<ul>\n<li><strong>Coding Agents<\/strong>: It recorded 61.7% on &#8220;DeepSWE v1.1&#8221; and 59.4% on &#8220;SWE-Atlas-QnA&#8221;. Additionally, it scored 28.3% on &#8220;Terminal-Bench 4&#8221; (Terminal-Bench 4.0), which measures whether an agent can actually execute commands in a terminal and complete assigned tasks. Their integrated score, the Coding Agent Index, reached 49.8%, showing performance superior to DeepSeek V4 Pro 0813 and Qwen3.8 Max.<\/li>\n<li><strong>Blind Evaluation<\/strong>: In a blind evaluation of coding quality (out of 5 points) conducted jointly with Surge AI by professional evaluators, the model scored 3.74, ranking second behind Kimi K3 (3.59) and GLM-5.3 (3.60) [Note: The source text comparison order implies relative placement, though Kimi\/GLM scores are noted as lower in the original]. However, it did not reach the top-ranked Claude Opus 5 (4.22).<\/li>\n<li><strong>Business Workflows<\/strong>: It recorded 59.9% on &#8220;AutomationBench&#8221;, which evaluates app-integration tasks, outperforming Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro.<\/li>\n<\/ul>\n<h3>Image Understanding and Visual Grounding<\/h3>\n<p>Performance exceeding closed models has been reported in multimodal performance combining images and text, particularly in visual grounding (detection and localization of specific objects within an image).<\/p>\n<ul>\n<li><strong>Dense 200<\/strong>: It recorded 42% in visual grounding tasks, showing results that exceed GPT-6 Astra&#8217;s 41%.<\/li>\n<li><strong>Chart Comprehension<\/strong>: It also demonstrates strong reasoning capabilities in tasks such as &#8220;ChartQA&#8221;, which reads numerical values and trends from graph images.<\/li>\n<\/ul>\n<h3>Comparison with Other Models<\/h3>\n<p>Data from the publishers and third-party reports show that this model has the capability to surpass or rival top-tier closed models and competing open models in cybersecurity, visual grounding, and specific STEM fields (CAD, physics, and mathematics). On the other hand, it falls slightly behind Claude Opus 5 in general coding quality blind evaluations, and some external benchmarks (such as the Vals Index) point out that this model (48.05%) scored lower than GLM-5.3 (53.51%). Thus, there is a trade-off where it exhibits very strong characteristics in specific specialized workloads while not dominating all comprehensive metrics.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>Mistral Large 4 demonstrates its true value in highly specialized practical applications by combining advanced reasoning capabilities and multimodal comprehension.<\/p>\n<h3>Cybersecurity<\/h3>\n<p>One of the model&#8217;s most prominent areas of expertise is cybersecurity. It is suitable for tasks such as actually reproducing software vulnerabilities and applying correction patches, malware analysis, vulnerability prioritization, and creating detection rules. While many closed models suffer from issues where legitimate vulnerability research and incident response requests are rejected due to safety filter limitations (provider-level refusals), deploying this open-weight model in an in-house environment allows organizations to execute advanced security operations unconstrained by external policies.<\/p>\n<h3>Advanced Agent Workflows<\/h3>\n<p>It is suitable for use as an agent that understands complex instructions and autonomously utilizes tools to complete tasks. It demonstrates high processing capabilities in automating business workflows linked with external applications such as Gmail, Google Sheets, Slack, and Salesforce, as well as creating and editing complex business documents such as spreadsheets, presentation slides, and PDFs.<\/p>\n<h3>Software Development and Coding<\/h3>\n<p>It excels in software engineering that understands the entire repository structure and workflows involving complex command operations in the terminal. It is envisioned for use as a development support agent that implements functions according to specifications and performs code debugging.<\/p>\n<h3>Visual Grounding and Document Analysis<\/h3>\n<p>Equipped with a 1.6 billion parameter vision encoder, its image understanding capability has improved dramatically. It can handle extremely precise visual grounding tasks, ranging from reading complex charts and graphs and extracting evidence data from PDFs, to analyzing gigapixel-class ultra-high-resolution satellite images or zooming in to inspect details of industrial blueprints. This strongly supports visual decision-making in fields such as disaster response, manufacturing, and engineering.<\/p>\n<h3>Science and Mathematics (STEM)<\/h3>\n<p>It excels in implementing specialized scientific workflows in physics, mathematics, materials science, and biology. It specializes in generating data analysis and simulation code, functioning as an advanced academic research assistant that can, for instance, generate complex chemistry &#8220;Hartree\u2013Fock simulation&#8221; code in a single shot.<\/p>\n<h3>Multilingual Business<\/h3>\n<p>It supports over 160 languages, covering all official languages of the European Union (EU). It can be utilized as a core model for global business expansion, including advanced multilingual document creation, translation, and legal or financial analysis tasks.<\/p>\n<h2>How to Get It<\/h2>\n<p>Mistral Large 4 is currently in a public preview phase.<\/p>\n<h3>Using the Preview API<\/h3>\n<p>The model can currently be tested via the preview API on Mistral AI&#8217;s platform, &#8220;Mistral Studio&#8221;. API pricing is set at $1.36 per million input tokens and $4.18 per million output tokens.<\/p>\n<h3>Planned Weight Release<\/h3>\n<p>Model weights for running the model by deploying it to local environments, private clouds, or on-premise infrastructure are scheduled to be released by the end of October 2026. Once the weights are released, organizations will be able to operate this model completely independently on their own infrastructure in compliance with their own internal security policies.<\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/mistral-large-4-released\/\">Mistral AI Releases Mistral Large 4 Open-Weight Model<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/mistral-ai-releases-timestral\/\">Mistral AI Releases Timestral: Time Series Foundation Models<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/mistral-ai-announces-voxtral-3\/\">Mistral AI Announces Voxtral 3 Audio Models and Voice Agent SDK<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/mistral-large-4-released-2\/\">Mistral Large 4 Released: Multimodal MoE Model<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>How to read Terminal-Bench, ChartQA<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-benchmarks-en\/\">Benchmark glossary<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/mistral.ai\/news\/mistral-large-4\/\">Introducing Mistral Large 4<\/a><\/li>\n<li><a href=\"https:\/\/docs.mistral.ai\/models\/mistral-large-4-0\">Mistral Large 4 Model Details<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Mistral AI has launched a public preview of Mistral Large 4, its flagship 1.05T parameter model with multimodal capabilities and strong cybersecurity.<\/p>\n","protected":false},"author":1,"featured_media":10348,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[310],"tags":[393,3075,165,1547,896],"class_list":["post-10349","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-new-models","tag-mistral-ai-en","tag-mistral-large-4-en","tag-moe-en","tag-verified","tag--en"],"lang":"en","translations":{"en":10349,"ja":10347},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/10349","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=10349"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/10349\/revisions"}],"predecessor-version":[{"id":17708,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/10349\/revisions\/17708"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/10348"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=10349"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=10349"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=10349"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}