Revolut has made GLM, an open-weights model from China’s Z.ai, the default coding assistant for about 2,000 engineers. David Abitbol, the company’s head of AI operations and infrastructure, said the fintech now reserves tools from Anthropic and OpenAI for its most difficult work.

Speaking at Nvidia’s GTC event in Berlin, Abitbol said Revolut pays roughly a tenth of the cost of comparable frontier models. The company runs GLM on its own servers, serving engineers with around 56 Nvidia GPUs. He put the combined cost at about 30 cents per million tokens, with a fixed bill rather than one that rises with each request.

Abitbol reported that the change reduced coding time by around 60% and the time from assigning an issue to deploying it by about 70%. He also outlined Revolut’s earlier steps: Google’s Gemini, introduced in January, improved delivery time by about 10% and coding time by 20%, according to him. The company later offered Claude Code with a monthly allowance of $300 per engineer, but Abitbol said the productivity gains were small because the allowance did not last the full month. The self-hosted model, introduced this year, came with unlimited tokens.

With token costs no longer tied to individual requests, engineers also used the system for work around coding, including searching Confluence, handling Jira issues and finding items in Slack. Abitbol said this highlighted how much of the team’s workload involved tasks beyond writing code.

The deployment required tuning. Abitbol said that in June, response times exceeded 10 seconds at four requests per second. After adjustments using the same number of GPUs, the system handled more than eight requests per second with latency below 1.8 seconds. His recommendations included limiting the context window to 256,000 tokens, using lower-precision FP4 rather than FP8 to save around 40% of GPU capacity, and monitoring cache hit rates instead of adding hardware whenever traffic rises.

Addressing the use of a Chinese model in a regulated financial company, Abitbol said Revolut uses only the model weights and hosts them on its own infrastructure or with a provider in the EU or US. He said no data is sent to a Chinese service. The agents run in a secure sandbox that permits some commands and blocks others, with a person remaining in the loop. GLM is text-only, so Revolut routes image tasks to a cheaper model such as Gemini Flash.

Sources: Xataka.