Cohere has uploaded the weights for Command A Plus to Hugging Face under the CohereLabs organization, making the updated enterprise model available for self-hosted deployment. The file is labeled command-a-plus-05-2026-bf16 - the bf16 suffix meaning it ships in bfloat16 format, a 16-bit floating-point precision that uses roughly half the memory of standard 32-bit weights while holding up better under training and inference than plain float16.
The original Command A launched earlier this year as Cohere's flagship model for enterprise agentic work - long document processing, multi-step reasoning, and tool use across business workflows. A "plus" revision typically signals improved benchmark performance or expanded capability, though Cohere hasn't published a technical report alongside this weights drop, so the specific improvements aren't yet documented.
For teams that care about this: Command A is one of the few serious enterprise models where you can actually run the full weights on your own infrastructure. Most Cohere customers use the API, but self-hosting bf16 weights matters for organizations with data residency requirements or high enough volume that per-token API costs become a real line item. The local LLM community on Hugging Face will benchmark this against the base Command A quickly - expect comparison runs within days.