Kimi K3 Max Claims Wins on Internal Agent Tests
The model reportedly beats Claude and GPT on real user tasks
@Kimi_Moonshot1 day ago
TL;DR:
- Kimi tested its latest model on actual jobs like experiments, slide decks, and finance work
- Results look good but stay private so hard to check
- Big labs seem more focused on agent reliability than benchmark scores now
Kimi's New Model Claims
Kimi.ai says its K3 Max model beats Claude Opus 4.8 and GPT-5.5 on some internal tests for agent workflows.
What They Tested
The company ran checks across three areas: online experiments, deck creation, and finance tasks. They argue the numbers reflect real gains in how agents handle practical work rather than just the usual public benchmarks.
The Private Benchmark Angle
The notable bit is the focus on benchmarks pulled from actual user jobs. If the results hold, the model could manage enterprise tasks like pulling research together or running financial checks. Without public data though, it's tough to verify. This also tracks with the big labs shifting toward reliable agent performance over pure reasoning scores.