Cohere Releases Command A+: A 218B Sparse MoE Model for Agentic Workflows That Runs on as Few as Two H100 GPUs

Cohere has unveiled Command A+, an open-source 218B sparse Mixture-of-Experts (MoE) model, specifically designed for enterprise agentic workflows. Released under an Apache 2.0 license, this model integrates functionalities from four prior models—Command A, Command A Reasoning, Command A Vision, and Command A Translate—into a single, scalable entity. It is optimized for diverse applications including reasoning, RAG, multilingual processing, and multimodal document handling, running efficiently on minimal GPU infrastructure.

Command A+ is a decoder-only Sparse MoE Transformer with 218 billion total parameters and 25 billion active parameters. It utilizes 128 experts, with eight active per token, alongside a single shared expert. This architecture routes each token through a subset of expert sub-networks, maintaining active compute at a 25B-parameter scale during inference, which significantly enhances performance while minimizing computational overhead.

Input modalities for Command A+ include text, image, and tool use, with output modalities encompassing text, reasoning, and tool use. The model supports an expansive 128K input context length and a 64K maximum generation length. Quantization variants, such as W4A4, enable the model to operate on a single B200 or two H100 GPUs, making it highly accessible for various deployment scenarios.

The model demonstrates substantial performance improvements across multiple benchmarks. On τ²-Bench Telecom, scores surged from 37% to 85% over Command A Reasoning, and Terminal-Bench Hard agentic coding performance reached 25% from 3%. Internal evaluations show a 20% accuracy improvement in Agentic Question Answering and a 32% rise in spreadsheet analysis quality. Command A+ also achieves 63% on MMMU Pro and 75.1% on MMMU, highlighting its multimodal reasoning capabilities.

Command A+ expands its multilingual coverage from 23 to 48 languages, yielding gains in machine translation and multilingual reasoning. Its tokenization efficiency has also improved, by 20% for Arabic, 16% for Korean, and 18% for Japanese, reducing the tokens needed for response generation. The model further offers up to 63% higher Output Tokens per Second and reduces Time To First Token by up to 17% compared to Command A Reasoning, along with an additional 1.5–1.6× inference speedup from speculative decoding.

Cohere employs NVFP4 W4A4 quantization for MoE experts and Quantization-Aware Distillation (QAD) to maintain high quality. Tool use is managed through Transformers chat templates using JSON schema, with reasoning traces available for detailed analysis. Command A+ is supported by vLLM and Transformers, with specific version requirements for optimal performance, and Cohere provides recommended sampling parameters for effective use.