Cohere releases Command A+, an open-source 218B MoE enterprise model that runs on two H100s and cuts agentic task latency by 30%

Cohere on May 20, 2026 released Command A+ (model ID: command-a-plus-05-2026), a 218-billion-parameter mixture-of-experts model under the Apache 2.0 license, targeting production enterprise deployments that demand strong agentic reasoning, vision, and multilingual capabilities without the infrastructure overhead of dense models.

What's new

Command A+ is Cohere's first MoE architecture at this scale. The headline specs:

• 218B total parameters, 25B active parameters — the MoE architecture activates only a fraction of weights per token, enabling high-capability outputs at reduced compute per inference step
• 128K input context, 64K max generation
• Apache 2.0 license — commercially permissive, full open weights
• Hardware floor: runs on as little as two NVIDIA H100 GPUs or a single NVIDIA Blackwell GPU
• 48 languages supported, including all official EU languages — more than double Command A Reasoning

Benchmark gains over Command A Reasoning:

• Tau2-Bench Telecom: 37% to 85%
• Terminal-Bench Hard (agentic coding): 3% to 25%
• Internal agentic Q&A: +20%
• Spreadsheet analysis: +32%

Throughput and latency: 63% higher output tokens per second versus Command A Reasoning; time-to-first-token reduced by up to 17%.

The model supports text, image, and tool use as inputs and is described by Cohere as "an open-source enterprise workhorse built for complex reasoning, multimodal and multilingual agentic tasks."

Availability: Hugging Face (BF16, FP8, and W4A4 quantizations), Cohere Model Vault for managed inference, and free trial via Hugging Face Space or the Cohere API.

Context

Command A+ completes Cohere's Command A family. The trajectory from Command A through Command A Reasoning to Command A+ reflects an explicit push toward enterprise-grade agentic performance — where tool use, multi-step reasoning, and multilingual fidelity matter more than raw benchmark scores on single-turn tasks.

The 218B total / 25B active MoE architecture is Cohere's answer to the efficiency problem: delivering close to dense-model-quality outputs while keeping hardware requirements low enough that a two-H100 deployment is viable for mid-sized enterprises, not just hyperscalers.

Why it matters

Open-weight MoE at this scale with a permissive license remains relatively uncommon. The combination of strong agentic benchmarks, EU-language coverage (positioning it clearly for European enterprise compliance scenarios), and modest hardware requirements makes Command A+ a credible alternative to proprietary API-only models for organizations with data-residency requirements. The Apache 2.0 license means enterprises can deploy, fine-tune, and modify the model without Cohere license restrictions — a meaningful differentiator in regulated industries.