Command A Plus

Cohere has open-sourced Command A+. Released under the Apache 2.0 license, this model adopts a mixture-of-experts (MoE) architecture, activating only 25B of its total 218B parameters, and is an efficient LLM optimized for agentic tasks.

Key Specifications and Features:

• Context: 128K input, 64K generation
• Modality: Supports text, images, and tool use
• Languages: Supports 48 languages (expanded from 23)
• Reasoning Capabilities: Multimodal reasoning, RAG, multilingual document processing
• Hardware: With W4A4 quantization, can run on 1 B200 or 2 H100s

Performance Improvements:

Command A+ integrates the capabilities of previous Command A series models (Command A, Command A Reasoning, Command A Vision, Command A Translate) into a single model, while improving performance across all areas.

• Agentic Tasks: τ²-Bench Telecom 37%→85%, Terminal-Bench Hard 3%→25%
• Multimodal: MMMU 75.1%, MMMU Pro 63%, MathVista 73.5%→80.6%, CharXiv 46.9%→52.7%
• North Integrated Evaluation: 20% improvement in agentic QA, 32% improvement in spreadsheet analysis, memory utilization 39%→54%
• Multilingual: Improvements in machine translation and multilingual reasoning, language support expanded from 23 to 48

Efficiency:

With its MoE structure, at the same quantization and concurrency levels, it achieves a 63% increase in Output Tokens per Second (TOPS) and a 17% reduction in Time To First Token (TTFT) compared to Command A Reasoning. W4A4 quantization provides an additional 47% speed improvement and 13% latency reduction, with almost no quality degradation.

The model is available on Hugging Face in BF16, FP8, and W4A4 quantized versions, and supports the vLLM and Transformers frameworks. It can also be deployed as a managed inference environment through Cohere's Model Vault.

On the Artificial Analysis Intelligence Index, it scored 37 points, demonstrating general-purpose enterprise agentic workflow performance that surpasses major open models.