Revolut has made a self-hosted version of Z.ai’s open-weight GLM model the default coding model for roughly 2,000 engineers, according to David Abitbol, the British fintech company’s head of AI operations. Speaking at the AI Engineer conference in New York, Abitbol said Revolut cut AI coding inference costs by about 90% in less than a year.
He said the rollout of AI coding tools improved coding time by 60% and lead time by 70% across that engineering group. Revolut said those savings came from deploying GLM models from Chinese AI company Z.ai on its own infrastructure.
Budget caps came before the move to self-hosting
Revolut introduced Google’s Gemini in early 2025 and Anthropic’s Claude Code in early 2026. Abitbol said, 「We were highly skeptical of AI costs. We saw engineers spending $1,000 a month, $10,000 a month, and the return on investment was not that high.」
That led the company to set strict budget caps when it rolled out AI coding tools. With token usage limited for engineers, Revolut began looking at running models itself. The company said self-hosted GLM-5.1 costs about $0.30 per million input or output tokens, which it described as roughly one-tenth the cost of OpenAI Codex or Anthropic Claude Code.
The presentation video did not say whether that figure includes GPU hardware, electricity, or operations staff.
Abitbol also said coding itself was not the biggest bottleneck. He said the larger gains came from using the model for debugging and for tickets in Jira and Confluence, Atlassian’s project management and documentation tools. 「When tokens are unlimited, you can really let go,」 he said. 「You don’t have to save tokens until the end of the quarter or reserve them for the hardest tasks.」
From GLM-5.1 to GLM-5.3 on 56 B200 GPUs
Revolut first used GLM-5.1 and later switched to GLM-5.3. It now runs the service on 56 Nvidia B200 GPUs. The Register calculated that this is equivalent to seven DGX B200 systems with eight GPUs each.
GLM-5.1 has about 754 billion total parameters, with roughly 40 billion activated for each generated token. Storing the model weights alone requires about 380 GB of GPU memory, and each user needs another 6 GB to 26 GB for context, depending on precision.
Revolut said its tuning work included model hyperparameters and NVFP4 quantization, which lowers numerical precision to reduce memory use and compute load. It also split prefill, the stage where the model reads the prompt, from decode, the stage where it generates output token by token, and adjusted context length. According to the presentation, those changes improved speed on the same hardware by 88%.
At around 256,000 tokens of context, the tuned system kept time-to-first-token below 3 seconds and generated about 200 tokens per second.
Vision tasks are routed to other models
GLM handles text only, which makes it a poor fit for interface design and other image-based work. According to The Register, Revolut uses model routing to send vision-related requests to Gemini or models such as GPT Luna, while keeping code generation on GLM.
Revolut said it recognizes the privacy and security concerns tied to using an open-weight model from China. The company’s view, as described in the report, is that the main concern lies in accessing such models through APIs hosted in China. If the model runs in Revolut’s own data center or in a regional cloud environment, it said there is little to worry about.
Key figures disclosed in the report
- About 2,000 engineers are using self-hosted GLM as the default coding model
- AI coding inference costs fell about 90% in less than a year
- Coding time improved 60%
- Lead time improved 70%
- Self-hosted GLM-5.1 costs about $0.30 per million tokens
- GLM-5.3 is now running on 56 Nvidia B200 GPUs
- Tuning improved speed on the same hardware by 88%
- At roughly 256,000 tokens of context, time-to-first-token stayed under 3 seconds and throughput was about 200 tokens per second


