c4ai-command-r7b-arabic-02-2025

C4AI Command R7B Arabic (c4ai-command-r7b-arabic-02-2025) is an open-weights 7-billion-parameter language model developed by Cohere and Cohere Labs, optimized for Modern Standard Arabic (MSA) alongside English. Built on the Command R7B foundation with continued pre-training on Arabic-centric data, supervised fine-tuning, and preference alignment, it targets enterprise use cases including instruction following, length control, retrieval-augmented generation (RAG), and culturally aware responses for the Middle East and North Africa region.

The model uses an optimized autoregressive Transformer architecture (Cohere2) with a 128,000-token context window, combining three sliding-window attention layers (window size 4096, RoPE) with a fourth global attention layer for efficient long-context modeling. It supports both conversational and instruct chat modes, multilingual RAG with inline citations via HuggingFace chat templates, and achieves best-in-class results among similarly sized models on Arabic benchmarks such as AlGhafa-Native (82.2), ArabicMMLU (60.9), IFEval Arabic (69.0), TyDiQA Arabic (83.0), and FaithEval Arabic (51.6), while retaining strong general capabilities on the HuggingFace Open LLM Leaderboard.

Released in February 2025 under CC-BY-NC 4.0 (with Cohere Labs Acceptable Use Policy), the model is designed for efficient local deployment—approximately 16GB VRAM in BF16 or roughly 5GB with Q4 quantization—making sovereign Arabic AI accessible on consumer hardware. Commercial use requires a separate license from Cohere; weights are available on HuggingFace and via Cohere's API and playground.