Cohere Labs' Command R7B Tops Open LLM Leaderboard Beating Gemma and Llama
Cohere Labs releases Command R7B, a 7B open-weights model that tops the HuggingFace leaderboard for its size class with first-class RAG, tool use, and agentic capabilities across 23 languages.
PRO- Command R7B is Cohere Labs' open-weights 7B model, topping the HuggingFace Open LLM Leaderboard for its size class with a 31.4 average score.
- Built on the new Cohere2 architecture, it mixes sliding window attention (4096 tokens) with periodic global attention layers for efficient 128K context handling.
- Natively supports RAG, multi-step tool use, and agentic loops via Transformers chat templates, with parallel tool calls allowed.
- Runs on consumer GPUs, CPUs, and MacBooks; available via Ollama, vLLM, and 26+ community quantizations.
- API pricing is $0.04/M input tokens and $0.15/M output tokens; open weights are free under CC-BY-NC.
- Covers 23 languages and excels at enterprise tasks: SQL, code translation, financial data retrieval, and HR/customer service summarization.
Command R7B is Cohere Labs' open-weights release of a 7-billion parameter model built squarely for enterprise workloads. It is the smallest and fastest entry in the Command R family, but it punches well above its weight class: it tops the HuggingFace Open LLM Leaderboard against similarly sized open-weights models, beating Gemma 2 9B, Ministral 8B, Llama 3.1 8B, and Qwen 2.5 7B on average score. The story here is not just benchmark numbers -- it is a model that was designed from the ground up to be deployed cheaply and integrated into real pipelines.
Built for the pipeline, not just the prompt
Command R7B is optimized for reasoning, summarization, question answering, and code, and is specifically trained to perform sophisticated tasks including RAG and tool use, with powerful agentic capabilities that let it use and combine multiple tools over multiple steps. That last part matters: most 7B models can call a single function. Command R7B is explicitly trained to chain tool calls across multiple reasoning steps, making it a viable backbone for lightweight agent loops.
It is the smallest and fastest model in the R family, offering state-of-the-art performance across real-world tasks and especially strong results for high-throughput, latency-sensitive applications like chatbots and code assistants. Its small size also unlocks dramatically cheaper deployment infrastructure -- including consumer GPUs and CPUs -- which means it can be used for on-device inference.
A hybrid attention architecture
The model runs on a new architecture called Cohere2, and its attention mechanism is worth understanding. Standard transformers apply full attention across every token in the sequence, which gets expensive fast. Command R7B takes a different approach: three out of every four layers use sliding window attention with a window size of 4096 tokens and RoPE (Rotary Position Embedding) for efficient local context modeling, while every fourth layer uses global attention without positional embeddings, enabling unrestricted token interactions across the entire sequence.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
