Canadian AI scaleup Cohere is joining the chorus of tech companies calling for AI models to be more transparent.
The news: In his very first post on X last week, Nvidia CEO Jensen Huang shared an open letter in support of releasing open-weight AI models to build a more competitive and accessible AI economy in the US. The letter defines open-weight models as large language models (LLMs) that can be freely downloaded, inspected, modified, and run on custom infrastructure. Initially launched with 25 signatories, including Nvidia, Meta, Microsoft, and Palantir, Huang’s letter amassed more than 50 signatures by Saturday. That’s when Cohere, the sole Canadian-headquartered company on the list and Canada’s only frontier LLM developer, announced it had signed on.
From the source: “This tech needs to be controlled by the people who use it, otherwise it’s just another piece of tech oligarchy control,” Cohere co-founder Nick Frosst wrote in an X post. “That’s why Cohere releases open-weight and open-source models. We will keep doing that.”
Following the thread:Since last week, nearly every big company in the AI space (including OpenAI, Google, and Canadian-founded, Palo Alto-based open-source AI startup Ollama) has signed the letter. The one major exception is Anthropic. That AI lab, which was last valued at $965 billion USD ($1.4 trillion CAD), appears to be sticking with closed-weight models, a month after the US government temporarily shut off access to its latest models for foreign nationals.
The letter comes as Chinese developers’ open-source AI models like DeepSeek and Kimi have exploded in popularity. The signatories warn against the “misappropriation” of US models, including through the technique known as distillation (using one model to train another), which the US Trump administration has accused Chinese companies of doing. The letter also calls for “targeted legal and commercial frameworks” to address the issue rather than put restrictions on open-source models.
Final thought: Non-American AI labs, like Cohere and France’s Mistral AI, have been more open-source friendly. With Frosst as a main champion, Cohere has made an intentional foray this year into open sourcing its AI models. It released its most powerful LLM so far, Command A+, as well as its speech-to-text transcription model, as open source. On The BetaKit Podcast, Frosst said part of the reason Cohere releases open-weight models is to empower users and to get the models in consumers’ hands, since the company mainly caters to enterprises with its flagship product, North.
Image courtsey The BetaKit Podcast.