Revolut cuts AI coding costs 90% by self-hosting GLM models

The company rolled out AI coding tools to roughly 2,000 engineers, reporting a 60 percent improvement in coding time and a 70 percent improvement in lead time. It started with GLM-5.1 and moved to GLM-5.3, both open weights models from Chinese developer Z.ai, served on 56 Nvidia B200 GPUs. Before that, Revolut had deployed Gemini in early 2025 and Claude Code in early 2026.

Cost was the driver. Abitbol said the company was seeing engineers spend $1,000 a month, $10,000 a month, with not so much return, so it launched with strict budget limits and limited tokens. Self-hosting GLM 5.1 at $0.30 per million input/output tokens cost about a tenth of using OpenAI's Codex or Anthropic's Claude Code, and cheap tokens meant engineers stopped rationing usage. Abitbol noted the biggest gains came from letting models handle debugging and Jira and Confluence tickets rather than code alone.

The hardware alone did not deliver the performance. Abitbol said the greatest gains came from tuning hyperparameters, switching to NVFP4 quantization, implementing disaggregated prefill and decode, and tuning context lengths. With those optimizations, Revolut achieves sub 3-second time to first token on roughly 256,000-token contexts and around 200 tokens per second of generation.

GLM has not replaced everything. The model is text only, so Revolut uses a model router to send vision queries to models such as Gemini or GPT Luna while code generation stays on GLM. Abitbol acknowledged that Chinese open weights models can raise privacy and security concerns, but argued the main risk is accessing them via APIs running out of China, saying there is little to worry about when running on-premises or on regional cloud resources.