Zed ships Zeta2.1, cuts prediction latency

Hot take: this is less a flashy model breakthrough than a strong systems win, and that is exactly the kind of improvement that compounds in a latency-sensitive feature like edit prediction.

• –The main value is efficiency: fewer output tokens directly reduces inference cost and response time.
• –A 28% faster p50 is meaningful for a keystroke-level product where users feel every millisecond.
• –Running 30% fewer servers for the same traffic is a strong signal that the release improves both UX and infra economics.
• –The “Multi-Region” format suggests the team found a better way to structure predictions, not just a better base model.
• –Keeping Zeta2.1 open-weight preserves the self-host and inspection angle that differentiates Zed’s model story.