Zeta2.1: 3x Fewer Tokens, 50ms Faster — Zed's Blog

Zeta2.1, Zed's open-weight edit prediction model, is now available with significant efficiency improvements over Zeta2. The update reduces average output tokens by 67% (~270 to ~90), cuts p50 response time from 189ms to 136ms, and requires 30% fewer servers for the same traffic. These gains come from a new 'Multi-Region' prompt format that outputs only the region around changed code rather than a large region around the cursor. Acceptance rate improved slightly (+0.51%) while explicit rejection rate dropped 4.1%. The model is open-weight on Hugging Face, trained on opt-in open-source data, and Rust prompt-formatting bindings are now published to PyPI for easier self-hosting.