c4ai-command-a-03-2025

Cohere Labs Command A (c4ai-command-a-03-2025) is an open-weights research release of a 111-billion-parameter dense decoder-only Transformer designed for enterprise workloads including agentic tool use, retrieval-augmented generation (RAG), multilingual chat, and code generation. Released in March 2025 under CC-BY-NC, it targets organizations that need high performance with modest hardware footprint—Cohere reports deployment on as few as two A100 or H100 GPUs with substantially higher throughput than the prior Command R+ refresh.

Architecturally, Command A uses grouped-query attention (GQA) with a hybrid attention pattern: three layers employ sliding-window attention (4,096-token window) with RoPE, while a fourth layer uses global attention without positional embeddings for full-sequence interactions. The model supports a 256K-token context window (128K default in the Hugging Face checkpoint), covers 23 languages, and is aligned via supervised fine-tuning and preference training for helpfulness and safety, with configurable contextual and strict safety modes.

On standard benchmarks, Command A is competitive with similarly sized and larger models: 85.5% MMLU, 69.6% MMLU-Pro, 50.51% GPQA, 90.9% IFEval, 80.0% MATH, and strong code results including 85.4% HumanEval (Python), 86.2% MBPP+, 26.9% LiveCodeBench, and 26.8% SWE-Bench Verified. It also excels at enterprise-oriented tasks such as SQL generation, multi-step tool calling with JSON-schema APIs, and RAG with optional citation grounding.