Revolut has developed GLM, China’s Z.ai’s flagship AI model, as the default programming assistant for its nearly 2,000 engineers. According to David Abitball, Revolut’s director of AI operations and infrastructure, anthropomorphic and OpenAI tools now remain for only the most complex tasks.
Local model vs frontier model. Local and self-hosted models have changed a lot in just two years. The most sophisticated ones can do things that previously seemed impossible, they’re cheaper to use and benefit you in addition to the fact that everything stays local. Revolut now relies on them in their daily lives, and many other companies are experimenting with this change.
Abitball told Nvidia’s GTC in Berlin that the Revolut now costs about a tenth of the price of comparable Frontier models. This is happening in a fintech with more than 80 million customers worldwide.
What has changed. As Abitbol says, Revolut engineers report on Anthropic’s Claude Code and OpenAI’s Codex. They now primarily work with GLM, an open weight model whose parameters are publicly available, and which Revolut runs on its own servers. Abitbol points out that frontier tools are still available, but as a reserved option for a small budget and the most difficult problems.
the numbers According to Abitbol’s discussion, Revolut used about 56 Nvidia GPUs to serve the model to its engineers. Their combined cost is about 30 cents per million tokens. The bill is also fixed without increasing with each request.
Results. Abitbol showed some information about this new change in strategy to about 2,000 engineers. The manager states that programming time is reduced by about 60% and delivery time by about 70% (from assigning an incident).
Model changes. Revolut first introduced Google’s Gemini in January, which Abitbol says has increased delivery time by about 10% and programming time by 20%. He then launched Claude Code with a budget of $300 per engineer per month. In this case, he says the increase in productivity was small because the speech didn’t last a full month. With the self-hosted model launched this year, the promise was unlimited tokens.
Does Revolut spend its tokens?. Abitbol noted that, without a budget to think about, they used tokens for tasks like searching Confluence (a business wiki), managing issues in Jira, or searching for items in Slack. As Abitbol points out, this shows that the real barrier wasn’t writing the code, but working around it.
the hard part Abitbol points out that the new configuration with the model does not work immediately. In fact, as he explained, in June, response times exceeded 10 seconds with four requests per second. After scaling the deployment to the same number of GPUs, Revolut was handling more than eight requests per second with a latency of less than 1.8 seconds.
Among the suggestions he made in his talk were limiting the context window to 256,000 tokens, saving about 40% of GPU power by using the lower precision FP4 format instead of FP8, and monitoring cache hit rates. Abitbol recommends resisting the temptation to add more hardware every time traffic increases.
The China question. Using a Chinese model in a regulated bank might raise some eyebrows, and Abitbol acknowledges this. But it highlights that Revolut only uses model weights and hosts them on its own infrastructure or with an EU or US provider, so no data goes to the Chinese service. He also claims it’s better for privacy than submitting code to OpenAI or Anthropy.
Additionally, Abilbot specifies that the agent runs in a secure environment (sandbox) that allows some commands and blocks others, always keeping a person in the loop. The GLM is text-only, so Revolut shifts imaging duties to cheaper models like the Gemini Flash.
Cover image appshunter.io
In Xataka | Anthropic’s IPO would achieve something extraordinary, with 80% of its employees becoming millionaires
