Cohere Targets Enterprise AI Translation with Command A Translate

On August 28, 2025, Cohere unveiled Command A Translate, a new large language model (LLM) built specifically for AI translation and positioned as “the new industry standard for secure, enterprise-ready […] translation.”

The company says it “outperforms all other leading models,” including GPT-5, DeepSeek-V3, DeepL Pro’s LLM, and Google Translate, across benchmarks in the 23 business languages the model is trained on.

Speaking to Slator, Tom Kocmi, Researcher at Cohere, said the company attributes a key gain to difficulty sampling — fine-tuning on the most challenging examples and down-weighting easier, lower-quality parallel data — to better target where the base model benefits most.

Cohere also adds Deep Translation, an agentic, multi-step refinement workflow designed to iteratively improve draft translations before delivery and boost quality on complex use cases.

“We are introducing a new agentic approach called Deep Translation, where the model leverages iterative reasoning for more complex translation use cases, and pushes performance even further above competition,” Jesse Willman, Staff Engineering Program Manager at Cohere, wrote in a LinkedIn post.

Cohere highlighted an external quality check: RWS Group has “confirmed Command A Translate’s strong performance across complex translation tasks” in multiple languages and domains, according to Cohere’s announcement.

“[Cohere’s] next challenge is to maintain that quality in the upcoming core general model release and enhance its overall multilingual capabilities.” — Tom Kocmi, Researcher, Cohere

‘Industry-Leading’ Performance, Full Data Control

The mode has 111B parameters with a 16k token context (8k input + 8k output) per the research model card, and — built on the Command A family — can run on one or two GPUs (A100s/H100s) for private deployments.

The launch taps a broader shift toward specialized, single-purpose models that can outperform much larger general-purpose LLMs.

Cohere is pitching the system to global enterprises handling “critical and sensitive” documents, emphasizing security and deployment flexibility. “Command A Translate delivers industry-leading performance while offering enterprises full control of their data through private deployment options,” Cohere said.

Command A Translate is available on the Cohere platform, can be run as a private deployment (for customers needing full tenant or on-prem control), and is accessible for research use on Hugging Face, with options for fine-tuning and customization to meet specific enterprise needs.

Kocmi noted in a LinkedIn post that the company is “releasing the weights so researchers can run wild with it,” and told everyone to wait for the technical report — speaking to Slator, he said it will be out late September.

Willman said he is “extremely excited about this model” and eager to see how customers adapt it for their localization and internationalization needs.

“Achieving top performance in AI translation is a critical milestone on the path to excelling in the multilingual space,” Kocmi told Slator, adding that their “next challenge is to maintain that quality in the upcoming core general model release and enhance its overall multilingual capabilities.”