{"id":16575,"date":"2025-12-04T16:35:53","date_gmt":"2025-12-04T15:35:53","guid":{"rendered":"https:\/\/haimagazine.com\/uncategorized\/mistral-3-unveiling-the-new-generation-of-open-weight-models\/"},"modified":"2025-12-08T12:08:45","modified_gmt":"2025-12-08T11:08:45","slug":"mistral-3-unveiling-the-new-generation-of-open-weight-models","status":"publish","type":"post","link":"https:\/\/haimagazine.com\/en\/ai-news-2\/mistral-3-unveiling-the-new-generation-of-open-weight-models\/","title":{"rendered":"\ud83d\udd12 Mistral 3: unveiling the new generation of open-weight models"},"content":{"rendered":"<p>Models with open weights are becoming increasingly competitive with closed commercial solutions. Mistral AI just unveiled the third generation of its products \u2014 the Mistral 3 family. Its launch is a step toward realizing the vision of distributed intelligence.<\/p><p>What does this mean in practice? Until now, to use the highest-quality AI, we typically had to connect to a powerful cloud server. The vision of distributed intelligence changes this approach. The idea is for AI to run where it&#8217;s needed, not only in a distant data center. With the different sizes of Mistral 3 models, developers can choose the right tool for the task: a gigantic model for complex analyses in the cloud, or a small, fast model that runs directly on your laptop, even without internet access.<\/p><h4 class=\"wp-block-heading\"><strong>Mistral Large 3 \u2013 a heavyweight in open architecture<\/strong><\/h4><p>The Mistral Large 3 model represents a significant technological leap that goes beyond standard updates. It&#8217;s based on the <em>Sparse Mixture-of-Experts<\/em> architecture. Although the total number of parameters is an impressive 675 billion, at any given time (for each token) only 41 billion are activated. As a result, the model retains extensive knowledge and capabilities while remaining surprisingly lightweight to operate.<\/p><p>The model was trained from scratch on a cluster of 3,000 state-of-the-art NVIDIA H200 GPUs. What are the results?<\/p><ul class=\"wp-block-list\"><li><strong>Performance:<\/strong> Mistral Large 3 achieves parity with the best instruction-tuned open-weight models on the market.<\/li>\n\n<li><strong>Rankings:<\/strong> In the prestigious LMArena list, it debuted in second place in the open\/source models category (excluding reasoning models) and sixth overall.<\/li>\n\n<li><strong>Versatility:<\/strong> he model not only generates text, but natively understands images and offers best-in-class support for languages other than English and Chinese.<\/li><\/ul><p>Each of them is available in three versions: base, instruct and reasoning.<\/p><h4 class=\"wp-block-heading\"><strong>Ministral 3 \u2013 &#8220;pocket-sized&#8221; intelligence with huge potential<\/strong><\/h4><p>R\u00f3wnie ciekawie jest na drugim biegunie skali. <\/p><p>Alongside the most powerful Large model, the Ministral 3 series debuted, dedicated to edge computing solutions. This term refers to processing data directly on the device (e.g., on a laptop or within local infrastructure) instead of sending it to external servers. In times when data privacy concerns are on the rise, the ability to run advanced AI locally is becoming crucial. The series includes three sizes:<\/p><ul class=\"wp-block-list\"><li>3 billion parameters (3B),<\/li>\n\n<li>8 billion parameters (8B),<\/li>\n\n<li>14 billion parameters (14B).<\/li><\/ul><p>Special attention should be paid to variants focused on reasoning. In situations that require ironclad logic, these models can &#8216;think&#8217; longer before providing an answer. This translates into impressive results \u2014 the 14B variant achieved as high as 85% accuracy on the AIME &#8217;25 math test.<\/p><p>Moreover, Ministral models are remarkably concise. Thanks to optimization, they often generate an order of magnitude fewer tokens than competitors while maintaining the same content quality. For businesses, this means significant savings on computational costs.<\/p><h4 class=\"wp-block-heading\"><strong>Technology alliance: NVIDIA, Red Hat and vLLM<\/strong><\/h4><p>The launch of Mistral 3 is also a showcase of hardware optimization, achieved through close collaboration with NVIDIA. All models were trained on the Hopper architecture, leveraging the bandwidth of HBM3e memory.<\/p><p>To ensure that model openness translates into practical utility, the engineering team focused on removing hardware barriers. Checkpoints were released in the new NVFP4 format, built using the llm-compressor tool.<\/p><ul class=\"wp-block-list\"><li><strong>For data centers:<\/strong> This allows for efficient operation of the Mistral Large 3 giant even on a single 8\u00d7A100 or 8\u00d7H100 node using vLLM.<\/li>\n\n<li><strong>For personal devices:<\/strong> Smaller models have been optimized to run on popular RTX cards, laptops, and Jetson platforms. This opens the door to implementing advanced AI in robotics, drones, or home assistants without the need to connect to the cloud.<\/li><\/ul><h4 class=\"wp-block-heading\"><strong>Availability and open license<\/strong><\/h4><p>In line with the company&#8217;s philosophy, the code and weights of the Mistral 3 family of models have been put online under the permissive Apache 2.0 license. This is an invitation for companies and developers to build their own solutions on a solid foundation, without fear of legal &#8220;vendor lock-in&#8221;.<\/p><p>Models are now available on leading platforms like Mistral AI Studio, Hugging Face, Amazon Bedrock, Azure Foundry, IBM WatsonX and OpenRouter. Soon, NVIDIA NIM and AWS SageMaker will join them. Organizations with specific needs can also use dedicated fine-tuning services for their own data.<\/p>","protected":false},"excerpt":{"rendered":"<p>Mistral AI is introducing the Mistral 3 model family, which includes the Large version with MoE architecture and the Ministral series for local devices. All models are available under the Apache 2.0 license.<\/p>\n","protected":false},"author":230,"featured_media":16515,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"rank_math_lock_modified_date":false,"footnotes":""},"categories":[813],"tags":[707,1004],"popular":[],"difficulty-level":[36],"ppma_author":[884],"class_list":["post-16575","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-news-2","tag-ai-5","tag-mistral-2","difficulty-level-easy"],"acf":[],"authors":[{"term_id":884,"user_id":230,"is_guest":0,"slug":"karolina-ceron","display_name":"Karolina Cero\u0144","avatar_url":"https:\/\/haimagazine.com\/wp-content\/uploads\/2025\/07\/PXL_20250419_110132091.MP4-scaled.jpg","first_name":"Karolina","last_name":"Cero\u0144","user_url":"","job_title":"","description":"Wsp\u00f3\u0142tw\u00f3rczyni newslettera AI Flash, studentka psychologii i pasjonatka sztucznej inteligencji. Interesuj\u0119 si\u0119 wp\u0142ywem nowych technologii na cz\u0142owieka, a w wolnych chwilach eksperymentuj\u0119 z generatywn\u0105 grafik\u0105 w Midjourney."}],"_links":{"self":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/16575","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/users\/230"}],"replies":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/comments?post=16575"}],"version-history":[{"count":1,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/16575\/revisions"}],"predecessor-version":[{"id":16576,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/16575\/revisions\/16576"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/media\/16515"}],"wp:attachment":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/media?parent=16575"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/categories?post=16575"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/tags?post=16575"},{"taxonomy":"popular","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/popular?post=16575"},{"taxonomy":"difficulty-level","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/difficulty-level?post=16575"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/ppma_author?post=16575"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}