• Sources: primary, discussion, discussion
  • Summary: Revolut ops and infrastructure lead David Abitbol told the AI Engineer conference that the company moved from Gemini and Claude Code to self-hosted GLM-5.1, then GLM-5.3, served on 56 Nvidia B200 GPUs, at about a tenth of the cost of Codex or Claude Code. He credits NVFP4 quantization, disaggregated prefill and decode, and a 256K context cap for sub-3-second time to first token and about 200 tokens per second. Revolut reports 60% faster coding time and 70% better lead time, with the largest gains from debugging and Jira and Confluence work.
  • Why it matters: It is a named-organization account of the tradeoffs in replacing frontier coding agents with open weights: hardware, serving tuning, latency targets and model routing.