Cohere has recently released its new language model, Command R Reasoning, designed to meet the stringent requirements of complex reasoning tasks in business environments. The model excels in handling agent workflows, building end-to-end systems, and analyzing large documents.
According to Cohere, Command R Reasoning performs exceptionally well in multiple industry benchmark tests, outperforming other private models, including BFCL-v3, Tau-bench, and DeepResearch Bench.
The model has powerful technical capabilities, capable of running on a single H100 or A100 GPU, supporting a context window of up to 128,000 tokens; with multi-GPU configurations, the context window can be scaled up to 256,000 tokens. Enterprises can flexibly set token budgets to effectively control costs and resource usage, enabling parallel processing of different workloads without switching models.
Command R Reasoning is particularly adept at handling complex research tasks. It can drive a hierarchical multi-agent system that breaks down complex requests into subtasks, performs parallel research, and ultimately generates structured reports. In relevant benchmark tests, this system achieved the highest scores in instruction following and content depth.
In terms of model availability, Command R Reasoning is currently available as a research version, with its weights open-source and licensed under the CC-BY-NC-4.0 license for free non-commercial use. For commercial deployment, enterprises need to contact Cohere's sales team.
Cohere places a high emphasis on model safety during development, striving to strike a balance between strict abuse prevention (covering child abuse, self-harm, violence, sexual content, and conspiracy theories) and avoiding excessive rejection of legitimate requests through an internal evaluation framework. Cohere states that compared to competitors, Command R Reasoning achieves the best balance between safety and practicality.
Currently, Command R Reasoning is officially available on the Cohere platform and Hugging Face.
