Cohere topped DeepL on WMT26 under a non-commercial licence

Cohere released North Small Translate, a mixture-of-experts translation model covering more than 50 languages that scores 83.60 on the WMT26 all-languages benchmark against 81.37 for DeepL NextGen and 68.20 for Google Translate. It ships under a CC BY-NC 4.0 licence, which means nobody can put it in a product.

Four items on today's page turn on permission rather than capability: who may use a model, whose catalogue it was built on, and who gets named when it produces something. The scores were the easy part.

The best number on the page cannot be sold

DeepL is the incumbent that benchmark is measured against, and its year has been hard on its own terms. It announced plans to cut about 250 employees, roughly 25% of its workforce, in May, and CEO Jarek Kutylowski said in June that the IPO market is not open to companies of DeepL's size. A rival scoring 2.23 points above it is a competitive event only if someone is allowed to ship the thing. Under CC BY-NC 4.0, nobody is.

Abacus.AI released three open-weight models on Hugging Face the same day: Smaug Agentic at 2T parameters based on Kimi K3, Smaug Flash fine-tuned on DeepSeek Flash, and a 27B Smaug Mini. The company claims its fine-tuning technique improves long-running agentic loops by 15 to 20% without increasing cost. The weights are open, and two of the three are built on somebody else's base model, so the terms a buyer actually has to read start upstream.

Buying the catalogue instead of arguing about it

ElevenLabs signed a multi-year licensing and joint product agreement with Universal Music Group, the first to cover both an AI audio company and a major label. It starts with a fan platform for remixes, mashups and personalized vocals built on licensed catalogue, and that platform is kept separate from the Music API and ElevenMusic products.

The separation is the part worth keeping. ElevenLabs crossed $500M ARR in May, up from $350M at the end of 2025, on products this deal does not touch. Licensed catalogue buys a new surface here; it does not retroactively clear the old one.

Credit is the other half of a licence

OpenAI withdrew its sponsorship of a mathematics event at CalTech on Thursday, days after 25 Fields Medal winners signed an open letter arguing that solutions announced in a rush leave no time for a proper writeup and raise severe attribution and plagiarism questions.

The money was never the issue. OpenAI closed $122B in committed capital at an $852B post-money valuation in March, which makes a conference sponsorship a rounding error it chose to stop paying. What the letter objected to is whose name goes on a result, and that is the one permission a cheque does not settle.

What an unrestricted release looks like

IBM and NASA released the open-source Lunar Foundation Model with a dataset of more than 30 spatially aligned layers from nine instruments across four missions. In the authors' own paper, it cut root mean square error by up to 22% against the SwinV2-B image model on predicting lunar ice potential.

It is the one release on today's page with no restriction on who may use it, and it came from a public agency and a vendor with nothing to protect in lunar ice.

A benchmark number says what a model can do. The licence says whether anyone will be allowed to do it. Today the highest score on the page carried the narrowest terms, and the widest terms came from a space agency.