Cohere

Contribution to AI

Two things were sold as one in the early years of commercial language models: the model, and the service wrapped around it. Cohere pulled them apart, and made the resulting pieces buyable by organisations that cannot send their data anywhere. Weights running inside a customer's own cloud, or on hardware behind a firewall with no outbound connection, turned banks, hospitals and defence ministries into plausible users rather than prospects waiting for a policy to change. The retrieval stack was unbundled in the same spirit: a cross-encoder that scores query-document pairs and reorders the output of whatever search system a company already runs, so improving search stopped requiring the replacement of search. The two Command R models of spring 2024 went further, training citation behaviour into the model itself — given a set of retrieved passages, the model marks which span of its answer came from which document, making attribution a property of generation rather than something bolted on afterwards. The larger of the pair, 104 billion parameters, was released with downloadable weights under a non-commercial licence and became briefly the highest-ranked non-proprietary model on the Chatbot Arena leaderboard, above one version of GPT-4. The non-profit lab set up in 2022 produced the more unusual result: Aya, instruction-tuned across 101 languages and released openly, built on annotations gathered from thousands of fluent speakers in languages that had almost no instruction data of their own.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).