Cohere Command R7B

Command R7B, published in December 2024, is described by Cohere itself as the smallest, fastest and final model of its R family. With roughly 8 billion parameters in the published weights and a 128,000-token context window, it was built for the part of the market the flagship models cannot serve: chatbots, code assistants and agents where latency and cost matter more than the last few points on a benchmark. The economics are the story. Cohere charges 0.0375 dollars per million input tokens and 0.15 dollars per million output tokens, which makes it roughly sixty-six times cheaper on input than Command A, the company's enterprise flagship. Cohere also points explicitly at consumer GPUs and CPUs, which puts on-device inference within reach — a deployment mode none of the larger Command models offer. The capability list is deliberately narrow but complete for enterprise work: retrieval-augmented generation with inline citations, native tool use, and multi-step agent loops. Cohere singles out the model's ability to avoid calling tools when they are not needed, which in practice matters as much as calling them well. The weights are on Hugging Face under a CC BY-NC 4.0 licence — free to download and inspect, but not to deploy commercially without a separate agreement with Cohere. That is the same arrangement as Command A and a different one from the Apache 2.0 licence the company later applied to Command A+.

▸Videos

No videos yet.

▸Reviews