TORONTO — Cohere has announced Command A, a new generative model the firm claims requires less processing power to produce results that rival leading products from OpenAI and DeepSeek.

The Toronto-based firm sells its technology to businesses, which use it to power applications or workplace productivity tools. “We focused on addressing some of the biggest pain points with enterprise AI adoption that I regularly hear from customers,” Cohere co-founder Nick Frosst said, citing hardware requirements, data security concerns and the cost of running models as key concerns.

As cloud providers and large firms compete for the chips to run AI applications, Frosst touted Command A’s efficiency. Cohere said its new model can run on just two of Nvidia’s in-demand A100 or H100 chips, compared to the dozens required by competitors. That makes it particularly useful to businesses that want to keep their AI tools on in-house hardware, for security or regulatory reasons.

Cohere also said Command A matches the performance of GPT-4o, OpenAI’s latest widely available model, and DeepSeek V3, the Chinese system that sparked a market rout earlier this year. Command A beat or matched GPT-4o and V3 on evaluations for academic knowledge, retail tasks like cancelling orders and changing addresses, and generating code. It also works faster, Cohere said. 

While some Silicon Valley firms are trying to improve their models using increasing amounts of information and compute, Cohere has focused on more modestly sized systems. But Command A “is not a super small model that gives up performance for efficiency,” Frosst said. It can do everything Cohere’s existing Command R+ product can, and “a lot more.” That includes handling inputs that are twice as long as its predecessor or GPT-4o, meaning that someone could, for example, drop in a lot more documents for the system to summarize.

Command A will be built into the startup’s North platform, which lets clients’ staff build AI agents to which they can delegate work. “We designed it to excel at real-world enterprise tasks [like] sourcing relevant HR policies or analyzing long financial reports with high accuracy,” said Frosst.        

Command A will be priced the same as Command R+, which is on par with what OpenAI charges for GPT4o. DeepSeek’s V3 costs about a tenth as much as the North American firms’ products.