**some conversation** Me: Let me check with the AI Gods. Wife: Which one? It's like Hinduism, there are so many.

am I the only one who remembers this banger looking at DeepSeek-OCR's DeepEncoder

There needs to be a voice agent that listens to people yapping and emails/pings others only if needed, thereby solving "This meeting should have been an email"

Grateful to have done so much with the most thoughtful and smart teammates I could have asked for. I'm sure we will continue to do great work - just keep shipping, no hype. Also personally wild to be part of a picture with a caption that has so many big beautiful numbers in it.

late night thoughts: > generational trauma is SFT > forging your own path is RL

Baffles me that in a city as beautiful as this, people still choose to be obsessed about SaaS and shit.

Read Story of Your Life and re-watched Arrival immediately - and this question blew my mind: Why did we not evolve written medium to be complementary to speech (and not the same)? Math expresses some aspects non-linearly and in less bits than natural language, but is still

Mistral Medium 3.1 is 2nd on LMArena without style control. Very proud of the @MistralAI team !

Introducing Mistral Medium 3.1. Overall performance boost, tone improvement, smarter web searches. Try it now in Le Chat (default model) or via our API (`mistral-medium-2508`).

In our continued commitment to open-science, we are releasing the Voxtral Technical Report: https://t.co/fIH9uW8qdZ The report covers details on pre-training, post-training, alignment and evaluations. We also present analysis on selecting the optimal model architecture, which

It's a great open-source model for transcription, but also most importantly: * You can use it as a single model for text-only applications (Voxtral-Small is just as good as Small 3.1) * Also supports goodies like native function calling, long-form audio summarization,

We also release a bunch of evals (e.g. GSM8K-speech, TriviaQA-Speech and MMLU-Speech) that we used internally to test trivia/math/general knowledge reasoning capabilities of our audio LLMs.

We did on-policy DPO to further improve the quality of model responses leveraging infrastructure built for Magistral and found that being on-policy helps a lot - especially to avoid regressions on ASR.

We do continued pretraining with two patterns: one to improve conversational/question-answering capabilities, and one to straight up maximize transcription performance. Balance of both is important to create an all-round amazing model.

The Voxtral tech-report is up! https://t.co/1qzV3ShvDE We release these models with a permissive Apache 2.0 license. Feedback is welcome! We have a lot more cooking, this is just the beginning.

Just a few months ago, I had no audio expertise beyond using it as a feature for video classification. Super proud of the first @MistralAI release in the audio space - it's been an amazing learning journey. There's of course more coming very soon!

Trying to get ahead of soham-gate: * I only work at @MistralAI * I'm very proud of the team I work with and the amazing models we ship. Reach out if you would like to do the same! https://t.co/6ZsB0xot1R

Tons of compute is not all you need - you need smart people who know how to use it efficiently. Come join us at @MistralAI and train the most cost efficient models!

Announcing Magistral, our first reasoning model designed to excel in domain-specific, transparent, and multilingual reasoning.

📰 News in Arena: Mistral Medium 3 makes a strong debut with the community! Highlights: 💠 #11 overall in chat: a +90 point leap from Mistral Large 💠Top-tier in technical domains (#5 in Math, #7 in Hard Prompts & Coding) 💠#9 in WebDev Arena Congrats to @MistralAI on the