Tip: Run Cohere Command A+ Locally at 4-bit (W4A4) With Lossless Quantization on a Single H100
VentureBeat·medium signal
Command A+'s W4A4 quantization format is notable because Cohere claims it's lossless — benchmark scores hold across BF16, FP8, and 4-bit formats. With only 25B parameters active per token (from the 218B MoE total), the 4-bit variant runs on a single H100 or equivalent. This is the first open-weight model at frontier capability that's specifically optimized for on-premises agentic deployment with native citation/grounding. For enterprises that can't send data to APIs, this is the new baseline local model for tool-use agents.