65.7% on Terminal-Bench 4.0 in Moonshot's own table, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra
Claim-level hallucination rate down from 23.4% on K3 to 8.3%
Cyber training data deliberately removed, which Moonshot says costs about nine points on xdev-bench-hard
Weights published on 4 November 2026 as promised, about 1.56 TB, under Moonshot's own licence rather than a standard one
A Model as a Service business above $20M in revenue needs a separate agreement, and the Kimi name must be shown on products above 100M monthly users or $20M monthly revenue. Internal use and Moonshot's certified partners are exempt
Best for long-horizon coding when the budget is the constraint.