# Revolut Says GLM Tuning Lifted Inference Speed 88% on 56 B200 GPUs By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/48761 Published: 2026-11-04T20:00:55.000Z Updated: 2026-11-04T20:00:55.000Z Section: Technology > Abitbol also cited GLM-5.1 at about $0.30 per million input or output tokens. Revolut AI operations lead David Abitbol said tuning raised inference speed by 88% on the same hardware, as the fintech uses its self-hosted model as the default coding model for about 2,000 engineers. Abitbol discussed the result at the AI Engineer conference in New York. Revolut moved from GLM-5.1 to GLM-5.3 and runs the current system on 56 Nvidia B200 GPUs. He described the 88% gain as an inference result but did not define the speed measure or identify a benchmark baseline. It does not establish that engineers completed coding tasks 88% faster. The tuning included changes to model settings and context length, as well as NVFP4 quantization, which reduces numerical precision to lower computing demands. Revolut also separated prompt processing, when the model reads an input, from answer generation, when it produces a response. In one configuration, GLM handled about 256,000 tokens of context, returned its first token in less than three seconds and generated about 200 tokens per second. Tokens are units of text processed by a model. The reported response time and generation rate describe model serving; Abitbol did not present them as a measure of how long engineers take to finish coding work. Abitbol cited a cost of about $0.30 per 1 million input or output tokens for GLM-5.1 and characterized it as roughly one-tenth the cost of OpenAI Codex or Anthropic Claude Code. Abitbol also said Revolut cut AI inference costs by about 90% in less than a year. The performance details add to Revolut’s earlier [move to make GLM-5.3 its default coding model for about 2,000 engineers](). Revolut began using Gemini in early 2025 and added Claude Code in early 2026. Abitbol said the company set strict token budgets before moving to self-hosting. Engineers now use GLM for debugging and work involving Jira and Confluence, as well as code generation. ## Links in this article - [move to make GLM-5.3 its default coding model for about 2,000 engineers](https://www.tokenpost.com/news/technology/47645)