Cohere releases Command A+: 218 billion parameters for agents on two GPUs

Cohere released Command A+ — an open-source model of the new generation for agentic workflows. This is 218 billion parameters in a Sparse Mixture-of-Experts architecture, combining four previous Command A variants into one universal solution.

Enormous Power in a Compact Form Factor

The main achievement is efficiency without loss of quality. Thanks to W4A4 quantization (4-bit recording of weights and activations), the model runs on two H100 GPUs. Competing models with 300+ billion parameters require eight to sixteen graphics cards. This reduces deployment costs by nearly four times. Support for 48 languages, including Russian, Chinese, and Arabic, makes Command A+ truly global. For companies building agents for international markets, this is critical. But what's most interesting is that this is Cohere's first multimodal reasoning model. It works with text, video, and images simultaneously, expanding the range of tasks from processing meeting recordings to analyzing screenshots and diagrams.

Who This Benefits

Open source code is a key advantage. Developers can now deploy Command A+ on their own servers without cloud dependency. For startups and companies with confidential data (fintech, healthcare), this is critical.

• Minimal requirements: two H100 instead of sixteen
• Multimodality in one model (text, video, images)
• Support for 48 languages for global markets
• Sparse MoE optimization: only 37B parameters work simultaneously
• Simplified lifecycle: four models became one

This is especially important for agents that require frequent updates and adaptation to specific workflows. When the model runs locally, the development cycle accelerates.

Context of Competition

Command A was previously released in four different variants — for search, chat, coding, and analytics. Combining into one multimodal model simplifies the ecosystem amid fierce competition with OpenAI, Anthropic, and other leaders. Cohere offers companies a powerful open-source foundation that researchers, startups, and enterprise clients can work with.

What This Means

Open large models are becoming more competitive with proprietary ones. When 218 billion parameters run on two graphics cards instead of a cluster of hundreds of GPUs, the barrier to entry drops sharply. For companies building their own agents, this means more control, lower costs, and a faster update cycle.