Reality Check:
- uses ~18.78% more tokens
- responds ~51.46% faster
- uses more haiku which is a smaller and less capable model than sonnet which makes the claim "3.2x better" questionable.
Just by instructing it to use haiku more.
@NyanpasuKA@chutes_ai It still happens, they did improve it but it's still way less stable than going to OpenAI, Gemini, Anthropic and having much better tool calling.
Claude Haiku 4.5 is out and it's bad. I don't know what they we're thinking but the only thing it learned from Sonnet 4 was to generate unnecessary files, not even close to Sonnet 4 performance on real projects.
KAT-Dev-72B-Exp๐ฅ @KwaiAICoder Kuaishou's new open model for software engineering
https://t.co/CknKPvNQNr
โจ 72B - Apache2.0
โจ Redesigned attention kernel & training engine for efficient context-aware RL
โจ 74.6% accuracy on SWE-Bench Verified