Can I run Command R 35B?

Command R 35B is Cohere's RAG-and-chat workhorse, and it asks more of your hardware than the 7-8B models most people start with. At 4-bit it weighs about 21 GB, so the 12 GB RTX 3060 is out entirely, and you need at least 32 GB of system RAM to load it comfortably. The realistic home for it is a 24 GB card like an RTX 4090, or an Apple Silicon Mac with plenty of unified memory. Drop to a 2-bit quant (about 15 GB) only if you are squeezed and willing to accept the quality hit.

In daily use it is steady rather than snappy. On an RTX 4090 you can expect around 40 tokens per second at 4-bit, which streams faster than you read; on an M-series Max it settles near 16 tok/s, usable but noticeably more deliberate; and on a CPU with DDR5 you are looking at roughly 2 tok/s, which is patience-testing territory. The 128K context window is genuinely large, but it is expensive to use: fill it and total memory climbs to about 54 GB, so unless you have a 48 GB-plus setup, keep working context modest and reserve the full window for occasional long documents.

Against the current crop it is showing its 2024 age. The Qwen 3.5 and 3.6 35B-A3B models in the related list are MoE designs that generally feel quicker and stronger on reasoning and coding, where Command R was never the focus. Its standout trait remains what Cohere built it for: grounded retrieval-augmented chat with citations, where it tends to stay on-source rather than wander. One hard caveat, though: the weights are CC-BY-NC, so this is non-commercial only. You can run it for personal projects and research, but you cannot ship it in a product or any paid service.