Revolut self-hosted GLM: 60-70% faster delivery at a tenth of frontier API costs
AI Engineer · youtube · 2026-10-29
David Abitbol, who leads AI operations at Revolut, shares data from 2,000 engineers on AI coding tools.
Key points:
- AI tools cut lead time across 2,000 engineers, but coding itself isn't the bottleneck
- Self-hosting open-source GLM models was the biggest win: 60-70% faster delivery at roughly a tenth of what frontier labs charged
- Tuning made the same GPUs 88% faster; he details the parameters that matter and the production setup
- Covers GPU requirements, latency targets engineers actually notice, and how an EU-regulated bank stays compliant
More from coding & agent
- Laravel Forge MCP Server lets Claude and Cursor manage servers, sites, and deployments — modelcontextprotocol · 2026-10-29
- Cohere Labs Open Science community hosts sessions on self-repairing training, agent harnessing, VLM safety — Cohere_Labs · 2026-10-29
- Housing Intel MCP chains FRED, BLS, ATTOM, and HUD APIs into one housing data source — modelcontextprotocol · 2026-10-29
- Indie dev open-sources Ask Your Data, a Text-to-SQL app that queries 1M-row loan data in plain English — Green_Highlight5508 · 2026-10-29
- 7 LLM eval tools compared: promptfoo, DeepEval, Ragas, Langfuse, Phoenix, LangSmith, Braintrust — ComfortableAbroad357 · 2026-10-29
- Unmix v0.2.0 adds post-edit checkpoints and a real Codex MCP pilot; Git still 27% faster — ashy2185 · 2026-10-29