Kimi 3 Trails Opus in Real-World Multi-Turn Tasks but Tops Benchmarks
Commentary on X
Spot on. Opus 4.8 just has a different level of understanding when it comes to holding context in multi-turn agentic loops. Still, we have to appreciate Kimi K3 taking the open-source crown! 🚀 Out of curiosity, what kind of specific use-cases were you testing them on? Do you think open-source is anywhere close to bridging this long-context gap soon?
@bindureddy For most things, Kimi K3 beats Opus 4.8
KIMI 3 IS NOT OPUS LEVEL - NOT IN THE REAL WORLD In our hard, multi-turn very long context evals, Kimi scores 50% of Opus 4.8 On first turn benchmark style coding problems, it actually beats Opus 4.8 🤷♀️