Cohere releases Command A+ under Apache 2.0: a 218B MoE model that runs on 2 H100s with 48-language support and 63% faster inference

Cohere released Command A+, its most capable model to date, on May 20, 2026 under an Apache 2.0 open-source license. The 218-billion-parameter Mixture-of-Experts model is designed to run on minimal hardware — a single NVIDIA B200 or two H100s — while delivering frontier-level performance on agentic tasks, multilingual workloads, and vision understanding.

What's new

Model: command-a-plus-05-2026

• Architecture: Sparse Mixture-of-Experts with 218B total parameters, 25B active per forward pass
• Context: 128K token input window, 64K max generation length
• Modalities: Text, image, and tool use — both input and output
• Languages: 48 languages, including all 24 official EU languages
• Hardware floor: 1× NVIDIA B200 (W4A4 quantization) or 2× NVIDIA H100s
• License: Apache 2.0 — fully permissive for commercial and private deployment
• Availability: Hugging Face, Cohere Model Vault, Cohere API (free trial)

Performance gains vs. Command A Reasoning:

• τ²-Bench Telecom (agentic benchmark): 37% → 85%
• Terminal-Bench Hard: 3% → 25%
• Agentic QA accuracy: +20%
• Spreadsheet analysis quality: +32%
• Inference speed: +63% output tokens per second
• Time to first token: −17% latency

The benchmark improvement on τ²-Bench Telecom — from 37% to 85% — is the most striking: it measures multi-step, real-environment agent tasks, and more than doubling the score represents a meaningful capability jump for enterprise agentic workflows.

Context

Cohere has historically positioned itself as the enterprise-first AI company: focused on private deployment, regulatory compliance, and on-premises or virtual private cloud infrastructure. Command A+ continues that strategy but adds something new — an Apache 2.0 license that makes it the company's first fully open-source model at this capability tier.

The hardware requirement is also notable in the open-model landscape. Most frontier-class open models (Llama 4, Qwen-A-72B, MiniMax M3) require multi-GPU clusters of eight or more H100s for comfortable serving. Running on 2× H100s puts Command A+ within reach of mid-scale enterprise infrastructure that doesn't have Blackwell or DGX systems.

The 48-language support, explicitly including all EU official languages, positions the model for European public sector and regulated-industry procurement — markets where French company Mistral has been dominant in the open-model category. Cohere's Canadian roots and compliance posture give it a different entry point to the same buyers.

Why it matters

For enterprises evaluating open-source LLMs, Command A+ changes the cost calculus at two levels: hardware and licensing. Apache 2.0 means no commercial restrictions or usage caps. The 2× H100 requirement means it can fit in an existing on-premises server rather than requiring a dedicated GPU cluster.

For the open-source AI ecosystem, it raises the performance bar for what runs on modest hardware. The agentic benchmark numbers — especially the near-tripling on τ²-Bench Telecom — suggest the model is genuinely competitive for tool-use and multi-step reasoning tasks that matter most for enterprise automation. Cohere's framing was direct: "Command A+ is an efficient, versatile, and privately deployable LLM built for high-performance agentic tasks with minimal compute overhead."