Cohere’s latest move is a study in efficiency without compromise. Command A+ packs 218 billion parameters into a sparse mixture of experts, and it runs on just two H100 GPUs. That’s a model that could fit in a modest server room, yet it outperforms its predecessor by leaps.
On τ²-Bench Telecom, scores jumped from 37% to 85%. Terminal-Bench Hard agentic coding? From 3% to 25%.
The secret lies in Quantization-Aware Distillation, a post‑training technique where the quantized student learns to mirror the full‑precision teacher’s output, using fake quantization on the forward pass and straight‑through estimators on the backward. The result: a model that closes quality gaps while staying lean enough for real‑world deployment.
To close residual quality gaps, Cohere uses Quantization-Aware Distillation (QAD) in the post-training phase: the quantized student model is trained to match the full-precision teacher’s output distribution, using fake quantization operators in the forward pass and straight-through estimators on the backward pass.
https://cohere.com/blog/command-a-plus
Performance vs. Prior Command A Models
On τ²-Bench Telecom, scores improved from 37% to 85% over Command A Reasoning, and Terminal-Bench Hard agentic coding performance reached 25% from 3%.
On internal North platform evaluations, all scored using LLM-as-a-judge techniques, Agentic Question Answering accuracy improved by 20% over Command A Reasoning.
This isn’t just a model release, it’s a redefinition of what’s possible on limited hardware. By compressing a 218B sparse MoE onto two H100 GPUs through Quantization-Aware Distillation, Cohere has turned a theoretical trade-off into a practical victory. The raw numbers tell the story: agentic coding jumps from 3% to 25%, telecom benchmarks leap 48 points, and QA accuracy climbs by a fifth.
Those aren’t incremental gains. They represent a structural shift in how frontier intelligence can be deployed, faster, leaner, and without sacrificing the fidelity that made the full-precision teacher worth training in the first place. Command A+ doesn’t ask for more compute.
It asks for a smarter approach. And that changes the conversation from “how much can we afford” to “what can we build today.”
How does Cohere's Command A+ 218B sparse MoE achieve such high performance on limited hardware?
Command A+ uses Quantization-Aware Distillation, a post-training technique where a quantized student model learns from the larger model to maintain performance while reducing size. This allows the 218 billion parameter sparse mixture of experts model to run efficiently on just two H100 GPUs without significant performance compromise.
What specific performance improvements does Command A+ show on τ²-Bench Telecom and Terminal-Bench Hard?
On τ²-Bench Telecom, Command A+ improved scores from 37% to 85%, representing a 48-point jump. For Terminal-Bench Hard agentic coding tasks, performance increased dramatically from 3% to 25%, demonstrating substantial gains in complex reasoning and coding capabilities.
What is the significance of running a 218B parameter model on two H100 GPUs?
Running such a large model on just two H100 GPUs represents a major efficiency breakthrough, as it could fit in a modest server room rather than requiring extensive data center infrastructure. This makes frontier-level AI capabilities accessible to organizations with limited hardware resources, redefining what's possible with constrained computational budgets.
How does sparse mixture of experts architecture contribute to Command A+ efficiency?
The sparse mixture of experts (MoE) architecture allows Command A+ to activate only a subset of its 218 billion parameters for each task, rather than using all parameters simultaneously. This selective activation, combined with Quantization-Aware Distillation, enables the model to maintain high performance while dramatically reducing memory and computational requirements.
We use cookies to analyze site traffic and improve your experience. By clicking "Accept", you consent to our use of cookies.
Learn more about our privacy policy
Open Source 
Open Source 
Open Source 
From Archives 
From Archives 