I'm joining SpaceX and xAI, working closely with Elon and team to build superintelligence.

Together SpaceX and xAI combine physical and digital intelligence under a leader who understands hardware at the deepest level. Add a high-agency culture with frontier-scale resources, and to achieve something truly unique.

I’m excited to advance the fields I’ve obsessed over for years, from robotics research to building AI models on the founding teams of Mistral and TML. Both were extraordinary journeys with extraordinary people that shaped how I think about building intelligence from the ground up.

Grateful for everything that brought me here and can’t wait to get started.

Career Update: Incredibly fortunate and excited to be part of the founding team at Thinking Machines Lab!

thinkingmachines.ai

Join us:
6wajk07p.paperform.co

What makes Le Chat unique?
- Lightning-fast responses: ~1000 words/sec!
- State-of-the-art document & image analysis
- Local code execution & data analysis
- Industry's leading image gen model

Lots of cool demos here: mistral.ai/en/news/all-ne…

Today, we are introducing the all new Le Chat: your ultimate AI sidekick for life and work! Now live on mobile!

Blog: mistral.ai/en/news/all-ne…

Try it on web: chat.mistral.ai

Download mobile apps here:
apps.apple.com/us/app/le-chat…
play.google.com/store/apps/det…

Performance of Mistral Small 3 Instruct model

Download on HF:
huggingface.co/mistralai/Mist…

4/N

Performance of Mistral Small 3 Base model

Download on HF: huggingface.co/mistralai/Mist…

3/N

Mistral Small 3 architecture is optimized for latency.

2/N

Announcing Mistral Small 3!
- 24B params, 81% MMLU
- Latency optimized: 150 tokens/s
- Competitive with Llama-3.3 70B, Qwen-2.5 32B, GPT4o-mini
- Apache 2.0

🔗Blog: mistral.ai/news/mistral-s…

1/N

We just released Codestral 25.01:
- Joint #1 on LMSys Copilot Arena while being much faster than other #1 models.
- 2x faster than Codestral 22B
- Strong FIM performance

Try it now for free in VSCode:
marketplace.visualstudio.com/items?itemName…

Blog: mistral.ai/news/codestral…

@MistralAI is growing! We opened a new office in Palo Alto, CA!

We are now hiring across research, engineering and business roles in US.

Join us: refer.hellotrusty.io/owkwk1hl1q

@nikshepsvn More details in the Pixtral Large blogpost:
mistral.ai/news/pixtral-l…

Le Chat gets new powers:
- Web search with citations
- Canvas for ideation, in-line editing, and export
- Vision: SoTA document & image understanding
- Image generation (Flux Pro)
- Speculative editing

And the best part - it's free!
Try it now: chat.mistral.ai

3/3

Pixtral Large:
- Frontier-class multimodal model
- SoTA on MathVista, DocVQA, VQAv2
- Maintains text performance of Mistral Large 2
- 123B decoder, 1B vision encoder
- 128K seq len

Download it on HF: huggingface.co/mistralai/Pixt…

Try it on Le Chat:
chat.mistral.ai

2/3

Today, we are announcing two new exciting updates:

Pixtral Large:
Frontier-class 124B multimodal model, powering the new Le Chat.

Brand new Le Chat:
With web search, canvas, image-gen, image understanding & more- all for free!

1/3

We just release two new models:
Ministral 3B and Ministral 8B

It’s been only a year since the release of Mistral 7B, and yet our smallest model today (Ministral 3B) already outperforms it on most benchmarks.

Blog: mistral.ai/news/ministrau…

We just released Pixtral 12B paper on Arxiv:
arxiv.org/abs/2410.07073

Strong performance by Pixtral 12B on the @lmarena_ai Vision Leaderboard!
- Best open-source Apache 2.0 model
- Outperforms Llama-3.2 90B while being 7x smaller
- Outperforms several closed models like Claude3 Opus

Try Pixtral for free on Le Chat: chat.mistral.ai!

Qualitative Examples:
Pixtral 12B can be used for:
- Reasoning over complex figures
- Chart understanding and analysis
- Multi-image instruction following
- Natural scene understanding
- Image to code
- Transcription of handwritten or scanned documents
and much more!

Performance vs open models
- Pixtral 12B outperforms all open models in its weight class at both text and multimodal
- Pixtral 12B particularly excels at Instruction following: substantially exceeds open models on Text and MM MT-Bench and IF-Eval benchmarks

Performance:
- All models re-evaluated with same evaluation harness
- Pixtral 12B outperforms several commercial closed models like Claude-3 Haiku and Gemini-1.5 Flash 8B
- Also outperforms much larger models like LLaVA OneVision 72B

Vision Encoder Architecture:
- New 400M vision encoder trained from scratch
- One image token for each 16x16 patch
- Image tokens flattened with [IMG BREAK] token between rows
- Allows variable image sizes & aspect ratios: High-res for diagrams & docs, low-res for icons/clipart

Architecture:
- Vision encoder tokenizes images
- Multimodal Transformer Decoder predicts the next text token given a seq of text and image tokens
- Pretrained with interleaved image-text data
- Allows any number of images in its large context window of 128K tokens

Super excited to announce Pixtral 12B - the first-ever multimodal model by @MistralAI!
- Strong on both text and multimodal benchmarks
- New 400M Vision Encoder
- Supports variable image sizes & multi-image input
- 128 seq len
- Apache 2.0

🔗 mistral.ai/news/pixtral-1…

Mistral Large 2 - Access and Customization
- Available today on La Plateforme and Le Chat
- Available on Google Vertex AI, Azure AI Studio, Amazon Bedrock & IBM Watsonx
- Also available for fine-tuning on La Plateforme along with Mistral Nemo 12B and Codestral 22B

6/N

Tool Use & Function Calling
Mistral Large 2 is equipped with enhanced function calling and retrieval skills and has undergone training to proficiently execute both parallel and sequential function call. It achieves state-of-the-art function calling performance.

5/N

Multilinguality
- Mistral Large 2 is trained on a large proportion of multilingual data.
- Excels in English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and Hindi.
- Outperforms Llama-3.1 70B, Comparable to Llama-3.1 405B

4/N

Instruction Following
Mistral Large 2 is particularly better at following precise instructions and handling long multi-turn conversations.
On Wild Bench, Arena Hard & MT Bench:
- Outperforms Llama 3.1 405B & Opus-3
- Comparable to Sonnet-3.5 and GPT-4o

3/N

Code & Reasoning
- Trained with 80+ programming languages
- Mistral Large 2 (123B) comparable to GPT-4o, Opus-3 and Llama-3 405B at coding benchmarks
- As compared to Mistral Large 1, significantly reduced hallucinations, improved reliability

2/N

Super excited to announce Mistral Large 2
- 123B params - fits on a single H100 node
- Natively Multilingual
- Strong code & reasoning
- SOTA function calling
- Open-weights for non-commercial usage

Blog: mistral.ai/news/mistral-l…
Weights: huggingface.co/mistralai/Mist…

1/N

Excited to announce Mistral-NeMo 12B trained in collab with @nvidia!
- Outperforms Gemma2 9B and Llama3 8B
- 128K context
- Multilingual in 100+ languages: excels in European, Asian & Indian languages
- Quant-Aware Training at FP8
- Apache 2.0

Blog: mistral.ai/news/mistral-n…

Excited to release Mathstral 7B:
- MATH performance: 56.6% pass@1, 68.4% with majority voting@64, 74.6% with a reward model
- Outperforms Minerva 540B on MATH
- 63.47% MMLU excelling in STEM
- Apache 2.0
- HF: huggingface.co/mistralai/math…
- Blog: mistral.ai/news/mathstral/

Excited to release Codestral-Mamba 7B:
- Infinite sequence length, tested with 256K tokens
- Linear time inference with Mamba
- Great for local code assistant
- Outperforms Transformers of similar size
- Apache 2.0
- HF: huggingface.co/mistralai/mamb…
- Blog: mistral.ai/news/codestral…

With $640M raise @ 6B+ valuation, Mistral AI has grown ~3x in valuation since December and ~25x in the last 12 months!

Join us to build the next generation of AI models in a small and extremely talented team. We are hiring in Paris, London & SF/Bay Area:
jobs.lever.co/mistral

Excite to announce: Mistral AI Fine-tuning API
- serverless fine-tuning on La Plateforme
- fast & cost-efficient using LoRA adapters
- easy deployment

Also, open-source library for fine-tuning on your own servers:
github.com/mistralai/mist…

Blog: mistral.ai/news/customiza…

Excited to announce our first code model: Codestral-22B!
- Fluent in 80+ programming languages
- Outperforms all open code models
- Available on our API and for free on Le Chat
- Integrated with VS Code
- Weights: huggingface.co/mistralai/Code…
- Blog: mistral.ai/news/codestral/

Mixtral-8x22B outperforms all open models across reasoning, knowledge, code and math capabilities while offering the best performance/cost ratio.

Mixtral-8x22B-v0.1 on HF:
huggingface.co/mistralai/Mixt…

Mixtral-8x22B-Instruct-v0.1 on HF: huggingface.co/mistralai/Mixt…

We just released Mixtral-8x22B-v0.1 and Mixtral-8x22B-Instruct-v0.1:
- Free to use under Apache 2.0 license
- Outperforms all open models
- Native function calling
- Masters English, French, Italian, German and Spanish.
- Seq_len = 64K

mistral.ai/news/mixtral-8…

Proud to announce: Mistral Large

- 81.2% MMLU
- outperforms Claude 2, Gemini 1.0 Pro, GPT-3.5
- multilingual with seq_len = 32K
- json mode and function calling

Also excited to launch Le Chat (Try it now at chat.mistral.ai)

🔗 mistral.ai/news/mistral-l…

Extremely proud of Mistral-Medium and Mixtral 8x7B performance on the @lmsysorg Leaderboard!

You can try both the models at console.mistral.ai

We just released Mixtral 8x7B paper on Arxiv:
arxiv.org/abs/2401.04088

Excited to release @MistralAI La Plateforme!

Three chat endpoints with competitive pricing:
Mistral-tiny: Mistral 7B Instruct v0.2, upgraded base model with higher context length 8K --> 32K and better finetuning, 6.84 --> 7.61 on MT Bench.
Mistral-small: Mistral 8x7B Instruct v0.1, 3.5 performance, multilingual.
Mistral-medium: Outperforms GPT-3.5 on all metrics, multilingual.

All endpoints have a 32K context size!

Also announcing Mistral-embed, an embedding model with a 1024 embedding dimension, achieves 55.26 on MTEB.

🔗 More details: mistral.ai/news/la-platef…
🔗 Register at: console.mistral.ai

Mixtral 8x7B is multilingual: vastly outperforms LLaMA 2 70B on French, Italian, German and Spanish with 6x faster inference

3/N

We also release Mixtral 8x7B Instruct v0.1 trained using supervised fine-tuning and direct preference optimisation (DPO).

It scores 8.3 on MT-Bench making it the best open-source model, with a performance comparable to GPT3.5.

huggingface.co/mistralai/Mixt…

2/N

Proud to announce:
Mixtral 8x7B -- Mixtral of Experts

- Free to use under Apache 2.0 license
- outperforms Llama 2 70B with 6x faster inference.
- matches or outperforms GPT3.5
- masters English, French, Italian, German and Spanish.
- seq_len = 32K

mistral.ai/news/mixtral-o…

1/N

GOAT: GO to AnyThing (Short Presentation)
A universal navigation system that can find and navigate to any object specified in any way - as an image, language, or a category - in completely unseen environments.

GOAT is result of several months of hard-work from a large team.

M.Chang, @theo_gervet , @mkulkhanna, @yvsriram, @shahdhruv_ ,@SoYeonTiffMin, K.Shah, @chris_j_paxton, @_saurabhg, @DhruvBatraDB, @RoozbehMottaghi, @JitendraMalikCV

Paper: arxiv.org/abs/2311.06430

GOAT also has a dynamic mapping capability which enables social navigation: find and follow people.

GOAT can readily be applied to downstream tasks such as pick and place.

GOAT improves with experience in the environment, from a 60% success rate at the first goal to a 90% success after exploration.

GOAT achieves an overall success rate of 83%, surpassing previous methods and ablations by 32% (absolute improvement).

In total, this represents over 200 different object instances given as image, language, or category goals.

We evaluate goat “in the wild” in 9 unseen rented homes without pre-computed maps or locations of objects. We evaluate 4 methods - GOAT and 3 baselines - for 10 trajectories per home with 5-10 goals per trajectory, for a total of 90 hours of experiments.

GOAT has a modular system design with a continually augmented Object Instance Memory that keeps track of the appearance of objects from different viewpoints. This enables navigation to targets specified by images and language descriptions.

GOAT has 3 key properties:

🖼️Multimodal: Objects can be specified in any way - image, language or category

🏃‍♂️Lifelong: It benefits from its past experience in the same environment

🤖Platform-Agnostic: It can be quickly deployed on robots with different embodiments.

Proud to present: GOAT: GO to AnyThing
A universal navigation system that can find any object specified in any way - as an image, language, or a category - in completely unseen environments.

Also useful for pick and place and social navigation!

theophilegervet.github.io/projects/goat/
🧵👇

Mistral 7B paper is on Arxiv:
arxiv.org/abs/2310.06825

Mistral 7B v0.1 is out!

- Best 7B model
- Outperforms Llama 2 13B on all benchmarks
- Outperforms LLaMA 1 34B on code, math, and reasoning
- Apache 2 license, no restrictions

mistral.ai/news/announcin…

Exciting personal update: After 5 years in FAIR and Meta AI, I have joined Mistral AI (@aimistral) with @arthurmensch, @GuillaumeLample and @tlacroix6 to build the next generation of AI!

For more details

PDF:
arxiv.org/abs/2212.00922

Webpage:
theophilegervet.github.io/projects/real-…

Presentation:
youtu.be/zZ6nlkTZVds

Led by @theo_gervet, w. @soumithchintala, @DhruvBatraDB, @JitendraMalikCV
(14/N)

Analysis: Why does modular learning work even better in reality than simulation?

Because failure modes in sim and reality have almost no overlap!

Takeaway: Navigation methods must be evaluated in the real world. (13/N)

The SemExp policy of the modular learning approach takes a semantic map as input, while the end-to-end policy directly operates on the RGB-D frames. The semantic map space is invariant between sim and reality, while the image space exhibits a large domain gap.
Example👇 (12/N)

Analysis: Why does Modular Learning transfer well while End-to-end does not?
To answer this question, we reconstructed one real-world home in simulation and conducted experiments with identical episodes in sim and reality. (11/N)

Key Result 3: End-to-end Learning Fails to Transfer
The end-to-end policy collides often, revisits the same places, and even fails to stop in front of goal objects when they are in sight.
Examples👇 (10/N)

Key Result 2: Modular Learning explores more efficiently than Classical methods
This is because Modular Learning is goal-oriented, it learns semantic exploration priors to find the goal object efficiently. Classical Frontier exploration is goal-agnostic.
Example👇 (9/N)

Key Result 1: Modular Learning is Reliable
We find that modular learning is very reliable on a robot, with a 90% success rate.
Examples👇 (8/N)