AI Agent Hub
Back to models
c4ai-command-r-08-2024 logo

c4ai-command-r-08-2024

Open Source CohereLabs Released 2024-08-30
-- 32.0B params 128K context Open Source

About this model

C4AI Command R 08-2024 is a 32-billion-parameter open-weight language model released by Cohere and Cohere Labs in August 2024 as a research refresh of the Command R family. Built on an optimized autoregressive Transformer with Grouped Query Attention (GQA), it supports a 128K-token context window and is trained on 23 languages while evaluated across 10 core languages. The model is post-trained with supervised fine-tuning and preference alignment for helpfulness and safety.

Command R 08-2024 is purpose-built for enterprise and research workflows centered on retrieval-augmented generation (RAG), grounded generation with inline citations, single-step function calling, and multi-step agentic tool use. It also handles reasoning, summarization, question answering, and code interaction, though it is optimized for code assistance rather than pure code completion. Compared with the prior Command R release, Cohere reports roughly 50% higher throughput, 20% lower latency, and about half the serving hardware footprint.

Weights are available on Hugging Face under a CC-BY-NC-4.0 license with Cohere Labs acceptable-use requirements, and the model is also served via the Cohere API as command-r-08-2024. Safety Modes (STRICT and CONTEXTUAL) were introduced in this refresh to give developers finer control over sensitive content handling in production deployments.

Benchmark Scores

GPQA
26.77

Technical Specs

  • Parameters: 32.0B
  • Architecture: Transformer
  • Context Window: 128,000 tokens
  • Input Modalities: text

Hardware Requirements

  • VRAM: 64.0 GB
  • Compute: Single NVIDIA A100 80GB for BF16 inference; 4-bit quantization fits on a 24GB GPU