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.

Running on 2 H100s is not “little hardware” . I got excited and then dejected

Our fastest, most powerful model yet. Command A+ combines high-performance agentic AI with efficient, practical deployment - running on as few as two H100s.

Faster: over 2x higher output speed and 30% lower latency than previous Command A models.

Smarter: across-the-board improvements for agentic, reasoning and multi-step tasks.

Worldlier: native support for 48 world languages and improved efficiency in non-European languages.

Releasing open-source under the Apache 2.0 license. We want to give developers direct access to enterprise-grade agentic capabilities from experimentation to production. Sovereign AI. For all. Download Command A+: Or learn more:

🤗

congrats on the release; where are the benches with other models?

Any open models are always good. But it would be great to compare against the SOTA open models (e.g. Qwen series). Also you state gpt-oss - which one? The big one or the small one?

Couple of compares i had generated from artificial analysis.

Canadian AI stays winning 🇨🇦🙌

@ClementDelangue can you also show comparison to models on the same size

Hmmmm I think I can actually w4a4 on 6 3090s going to try!

the fact its open weight is really cool nice work!

Two H100s for enterprise-grade agentic AI is the real headline. Compute arbitrage is where the next wave of enterprise adoption happens. Whoever solves sovereign deployment at this efficiency wins the market that GPT-4 can never touch.

I need more Vram I need more Vram I need more Vram I need more Vram

what is the point of this model? Qwen3.6-35B-A3B outperforms it on every overlapping bench. Terminal-Bench Hard 25% vs 34.8% MMMU-Pro 63% vs 75.3% CharXiv Reasoning 52.7% vs 78% τ²/TAU telecom 85% vs 95.3% And it is 218B-A25B vs 35B-A3B...

Video is gorgeous 😍

okieee, if this is better than 3.1 flash lite thats kinda huge for fast agent based products. need to try this in real world scenarios

Little hardware? 2 h100s isn’t little hardware and y’all aren’t beating qwen3.6 27b in benchmarks I see.

“Optimized to run on as little hardware as possible” might be the most important line here. Not everyone has a datacenter in their bedroom 😭 If startups, students, and indie devs can run powerful models cheaply, that changes who gets to build in AI.

👀

Running on just 2 H100s is the real story here. Most teams can't even get access to 8x setups. What's the actual tokens/sec throughput vs Llama 4 Scout on equivalent hardware?

This is a strong release for teams thinking about private agent deployments. Open weights plus long context, tool use, RAG, and efficient quantized serving makes the self-hosted agent stack feel much more practical. Congrats to the team.

'To all' Ah, yes let me pick out my two H100s just for this.

let me just check the couch cushions for two spare h100s real quick

@Thom_Wolf Haha 2x H100. I thought I can run on my mac mini M4 Pro with 24 GB

@huggingface oh nice, sounds like big efficiency gains

@jun_song

Comparing against your other models doesn’t give anyone any sense of how effective your model is.

🙏🏼🙏🏼🙏🏼

@0xAfterSnow

Awesome release 🇨🇦

Hey @Cohere team! I'm a developer from Brazil and I'm deeply interested in mastering Cohere, especially for RAG implementations. Could you look into providing Portuguese (PT-BR) audio translations or native subtitles for your YouTube and LLM University videos? The Brazilian AI

Powerful, low hardware, open source, Cohere really said take all three 🔥

This highlights how amazing qwen 3.6 and Gemma 4 are. It's hard to belive that the future hope of local (free)AI is in the hands of the Chinese. It's a wild word!!

@ClementDelangue @llmgram

@ClementDelangue Can you guys start showing these models being used for real enterprise use cases?

any more information about attention layers? 128k context is... not that much tbh