Cohere Command A+

Cohere Command A+ is the largest open-weight model in our catalog and the newest. It is a sparse mixture-of-experts model with 218B total parameters and 25B active per token, released by Cohere Labs under Apache 2.0 — so unlike Command A (CC-BY-NC, research-only), Command A+ is licensed for commercial use without a separate agreement.

The model is positioned by Cohere as their unified flagship: agentic-tuned (big jump on Terminal-Bench Hard), multilingual across 48 languages, and multimodal-capable. It is designed for enterprise workflows where one model needs to handle reasoning, tool use, RAG, document processing, and cross-language work in a single deployment.

When to pick it

• Enterprise agents that need strong tool use and multilingual reach in a single model — 48 languages including major European, Asian, and Middle-Eastern ones
• Long-context reasoning where 128K is enough but you want the strongest open-weight model in that window
• Workflows that mix code, language, and document understanding — the MoE architecture handles all three well
• When the procurement requirement is “Apache 2.0 or compatible” and you cannot use a CC-BY-NC research-only model

When to look elsewhere

• Very long context > 128K → Qwen 3.5 122B A10B (256K) or MiniMax M2.7 (1M)
• Latency-sensitive chat loops → DeepSeek V4 Flash or Gemma 4 26B A4B
• Pure Nepali-language workloads → HimalayaGPT 0.5B (free)

Hosting notes

We host Command A+ in W4A4-quantized form on a single Blackwell or a 2× H100 node depending on demand. Throughput characteristics will be confirmed publicly in the launch latency dashboard.

Credit

The model and weights are by Cohere / Cohere Labs. Model card and license on the Cohere blog and the Cohere Labs Hugging Face org.