Revolut Makes Z.ai's GLM Its Default Coding AI for 2,000 Engineers

The numbers

According to the talk, Revolut serves the model on around 56 Nvidia GPUs, at a combined cost of about 30 cents per million tokens. The bill is fixed rather than growing with each request. Abitbol said coding time fell by around 60% and delivery time, from ticket assignment to deployment, by around 70%.

The path was gradual. Gemini, introduced in January, improved delivery time by about 10% and coding time by 20%. Claude Code followed with a $300 monthly budget per engineer, but the gain was small because the allowance did not last the month. The self-hosted model promised unlimited tokens, and engineers used them for searching Confluence, handling Jira tickets and finding items in Slack. For Abitbol, that shows the real bottleneck was the work around writing code.

Not instant, and not without questions

The rollout was rocky at first: in June, response times exceeded 10 seconds at four requests per second. After tuning on the same GPUs, Revolut handled more than eight requests per second with latency under 1.8 seconds. His tips include capping the context window at 256,000 tokens, using lower-precision FP4 instead of FP8 to save about 40% of GPU capacity, watching cache hit rates, and resisting the urge to add hardware whenever traffic grows.

Using a Chinese model at a regulated bank may raise eyebrows, and Abitbol acknowledged it. He stressed that Revolut only uses the model weights, hosted on its own infrastructure or with an EU or US provider, so no data goes to a Chinese service. He argued this is even better for privacy than sending code to OpenAI or Anthropic. Agents run in a sandbox that allows some commands and blocks others, with a human always in the loop. Because GLM is text-only, image tasks go to a cheaper model such as Gemini Flash.