Connect with us

Hi, what are you looking for?

Top Stories

Mistral Launches AI Family with 675B Parameters, Competes Head-to-Head with DeepSeek

French startup Mistral launches four AI models, including the flagship Large 3 with 675 billion parameters, challenging DeepSeek’s dominance in the open-source arena.

French AI startup Mistral made a significant leap in the competitive landscape of artificial intelligence with the release of its latest model family on Tuesday. Known for its role as the underdog in a field largely dominated by American and Chinese firms, Mistral’s new offerings are positioned to challenge existing open-source models, providing them free of charge.

The new lineup includes four models, ranging from compact personal assistants to a cutting-edge system boasting 675 billion parameters. All models are available under the permissive Apache 2.0 open-source license, allowing users to download, modify, and fine-tune them for various applications on compatible hardware.

At the forefront is the flagship model, Mistral Large 3, which utilizes a sparse Mixture-of-Experts architecture, activating only 41 billion parameters for each token processed. This engineering choice allows the model to achieve performance comparable to much larger systems while operating at a level typically associated with 40 billion parameter models.

Trained from scratch using 3,000 NVIDIA H200 GPUs, Mistral Large 3 debuted impressively, ranking second among open-source, non-reasoning models on the LMArena leaderboard. In terms of benchmark comparisons, Mistral’s leading model surpasses DeepSeek V3.1 across several metrics but trails slightly behind the newer V3.2 version.

When it comes to general knowledge and expert reasoning tasks, Mistral’s offerings hold their ground, although DeepSeek maintains an edge in coding speed and mathematical logic. Notably, this new release does not incorporate reasoning models, which limits its cognitive capabilities compared to competitors.

The smaller models in the lineup, referred to as “Ministral,” are particularly noteworthy for developers. Available in three sizes—3 billion, 8 billion, and 14 billion parameters—these models come with both base and instruct variants and support native vision input. The 3B model has garnered attention from AI researcher Simon Willison, who highlighted its capability to run entirely within a browser using WebGPU.

This capability offers unique opportunities for developers and hobbyists alike, making it suitable for applications in drones, robots, and even offline systems in vehicles. Early testing has revealed a distinctive character across the Mistral lineup; the Mistral 3 Large demonstrates conversational fluency, often mirroring the style of GPT-5 but with a more natural cadence.

However, it has also shown a tendency for repetition and overreliance on common phrases, especially in its 14B instruct variant, which users have flagged on platforms like Reddit. Despite these issues, its ability to generate long-form content remains a highlight for its size.

The smaller 3B and 8B models, while functional, sometimes produce formulaic outputs on creative tasks, although their compact size allows them to run on less powerful hardware, such as smartphones. The only other competitive option in this niche is Google’s smallest version of Gemma 3.

Enterprise interest in Mistral is already materializing, as demonstrated by HSBC‘s announcement of a multi-year partnership to implement generative AI within its operations. The bank plans to self-host the models on its infrastructure, aligning Mistral’s expertise with its internal technical capabilities—a choice particularly appealing for organizations managing sensitive customer data.

In collaboration with NVIDIA, Mistral has developed a compressed checkpoint called NVFP4 that enables Mistral Large 3 to operate on a single node powered by eight high-end NVIDIA cards. NVIDIA claims that the Ministral 3B model achieves approximately 385 tokens per second on an RTX 5090, and around 50 tokens per second on Jetson Thor for robotics applications, highlighting its efficiency and speed without compromising quality.

Future developments include a reasoning-optimized version of Large 3, although competitors like DeepSeek R1 and various Chinese models retain their advantages in explicit reasoning tasks for now. For enterprises prioritizing cutting-edge capabilities, open-source flexibility, multilingual support, and compliance with European regulations, Mistral’s emergence marks a pivotal expansion of options in the AI landscape.

See also
Staff
Written By

The AiPressa Staff team brings you comprehensive coverage of the artificial intelligence industry, including breaking news, research developments, business trends, and policy updates. Our mission is to keep you informed about the rapidly evolving world of AI technology.

You May Also Like

Top Stories

OpenAI alerts U.S. lawmakers about Chinese startup DeepSeek’s suspected replication of American AI models, raising concerns over technology theft and security.

Top Stories

ByteDance's Seedance 2.0 generates high-quality videos mimicking Hollywood scenes, raising concerns over copyright and the future of traditional filmmaking.

Top Stories

US shares dropped 1.4% amid AI concerns while Australian stocks surged 2.4% on profit growth, signaling a shift from tech-heavy investments.

AI Technology

MiniMax launches the M2.5, achieving 100 TPS and transforming AI deployment costs to $0.3 input and $2.4 output per million tokens, enhancing operational efficiency.

Top Stories

Anthropic denies military use of its AI system Claude amid Pentagon tensions over a potential $200M contract and ethical concerns regarding autonomy and surveillance.

AI Generative

TikTok's Seedance 2.0, an AI video creation tool, sparks Hollywood backlash as users generate viral content featuring A-list stars like Tom Cruise and Brad...

Top Stories

OpenAI warns U.S. lawmakers that Chinese startup DeepSeek is allegedly cloning its ChatGPT models, raising national security concerns over AI technology theft.

AI Technology

DOE launches 26 AI challenges to cut nuclear deployment timelines by 50% and reduce operational costs by over 50% in a revolutionary energy initiative.

© 2025 AIPressa · Part of Buzzora Media · All rights reserved. This website provides general news and educational content for informational purposes only. While we strive for accuracy, we do not guarantee the completeness or reliability of the information presented. The content should not be considered professional advice of any kind. Readers are encouraged to verify facts and consult appropriate experts when needed. We are not responsible for any loss or inconvenience resulting from the use of information on this site. Some images used on this website are generated with artificial intelligence and are illustrative in nature. They may not accurately represent the products, people, or events described in the articles.