Roman Castagné

Member of Technical Staff @cohere, previously @InriaParisNLP @Inria, Ecole des Ponts ParisTech, MVA at ENS Paris-Saclay, intern at @NaverLabsEurope

[Roman Castagné @RomCast_]

This was a super fun project we did with @nthngdy, trying to keep the best of both worlds between tokenization-free models and subword tokenizers.

Check it out!

[Nathan Godey @nthngdy]

💥 What if you could train and fine-tune the tokenizer with your LM?
We introduce a new neural tokenizer named MANTa.
Amazingly, it learns to segment words w/o explicit supervision!

📜 Preprint: https://t.co/LoxYce2Vlv
Accepted at #EMNLP Findings

Thread below ⬇️ (1/8)

[Cohere @cohere]

Introducing Cohere's first open-source coding model: North Mini Code

Small & efficient, designed for agentic performance and built for community input.

[Cohere @cohere]

Introducing: Cohere Command A+

We’ve created our most powerful LLM yet, optimized it to run on as little hardware as possible, and released it open-source for all.

[Cohere @cohere]

Introducing: Cohere Transcribe – a new state-of-the-art in open source speech recognition.

[Cohere @cohere]

Introducing Command A Reasoning, our most advanced model for enterprise reasoning tasks.

[Cohere @cohere]

Introducing Command A Vision, a state-of-the-art generative model that excels across multimodal image capabilities that matter for enterprises!

[Roman Castagné @RomCast_]

@awnihannun @johnowhitaker Same thing for Llamba actually, they use the MLP weights, embeddings and layernorms from the teacher.

[Nathan Godey @nthngdy]

🚀 New Paper Alert! 🚀

We introduce Q-Filters, a training-free method for efficient KV Cache compression!

It is compatible with FlashAttention and can compress along generation which is particularly useful for reasoning models ⚡

⬇️R1-Distill-Llama-8B with 128 KV pairs ⬇️

🧵

[Aidan Gomez @aidangomez]

Today, we’re excited to release Command R7B, the smallest, fastest, and final model in the R series.

It’s a powerhouse with a minimal footprint.

[Cohere @cohere]

Introducing Command R7B: the smallest, fastest, and final model in our R series of enterprise-focused LLMs!

It delivers a powerful combination of state-of-the-art performance in its class and efficiency to lower the cost of building AI applications.

https://t.co/e2ah5c5x5J

[Cohere Labs @Cohere_Labs]

Aya Expanse: Combining Research Breakthroughs for a New Multilingual Frontier 🌿

Today, we release a technical report with extended evaluation for Aya Expanse, our new generation of 8B and 32B parameters multilingual language models 🌎

[Ekagra Ranjan @EkagraRanjan]

Building artefacts for JSON Schema is expensive to do on scale in production for LLM inference. Probably why OpenAI had "JSON mode" for a year but not "JSON schema mode" until @cohere released it first :)
* @cohere is >10x faster than OpenAI
* @cohere is >45x faster than Outlines

[Cohere @cohere]

We’re excited to announce our Series D financing to accelerate growth, expand our team, & develop our next class of frontier, enterprise-grade, data privacy-focused AI technology.

We’re bringing highly scalable and secure AI solutions to global enterprises wherever their data is and in whatever language they speak.

[Nick Frosst @nickfrosst]

@cohere just shipped json schema sampling!

Now not only can you guarantee that the model returns valid json, you can actually ensure it returns json with a specific format!

Big win for people actually building with LLMs :)

https://t.co/fUEEqZxTAw

[Arena.ai @arena]

Exciting news - the latest Arena result are out!

@cohere's Command R+ has climbed to the 6th spot, matching GPT-4-0314 level by 13K+ human votes! It's undoubtedly the **best** open model on the leaderboard now🔥

Big congrats to @cohere's incredible work & valuable contribution to the open community!

More exciting updates:
- Qwen1.5-32B-Chat almost top-10
- Gemma-1.1-7B-it shows great improvement (1044 -> 1088, on par with Llama-2-70b)
- Starling-7B-Beta still the best 7B with over 13K votes!

[Stella Biderman @BlancheMinerva]

Hot damn, @cohere @CohereForAI is really bringing it

[Aidan Gomez @aidangomez]

One subtlety worth mentioning is how significant the tokenizer is to the cost to use models in non-english languages. Our tokenizer is meaningfully better than others at the 9 non-English languages, achieving up to a 2x effective cost reduction to use.

[Cohere @cohere]

Today, we’re introducing Command R+: a state-of-the-art RAG-optimized LLM designed to tackle enterprise-grade workloads and speak the languages of global business.

Our R-series model family is now available on Microsoft Azure, and coming soon to additional cloud providers.

[Cohere Labs @Cohere_Labs]

Announcing C4AI Command R+ open weights, a state-of-the-art 104B LLM with RAG, tooling and multilingual in 10 languages. 

This release builds on our 35B and is a part of our commitment to make AI breakthroughs accessible to the research community. 🎉

https://t.co/2UCLl5sfPB

[Aidan Gomez @aidangomez]

⌘R+

Welcoming Command R+, our latest model focused on scalability, RAG, and Tool Use. Like last time, we're releasing the weights for research use, we hope they're useful to everyone! https://t.co/HgESxxEYlK

[Aidan Gomez @aidangomez]

Here are the weights: https://t.co/dr6SLoHT1s