afaik the Kimi-K2 team was the first to release a tool use training strategy at large scale following this a bunch of the other models started getting very good at tool use ie. GLM, Qwen etc.

I think about this every day: “At every job you should either learn or earn. Either is fine. Both is best. But if it’s neither, quit.” - Garry Tan

excited to be partnering with amazing folks Together AI, Zain and Ivan Provilkov to bring dynamic agent simulations to together evals.

📄New Guide: Running nanochat on instant clusters! Train and inference Andrej Karpathy's end-to-end ChatGPT clone on Together’s on-demand GPU clusters — and learn how to: ➡️Train nanochat ➡️Nanochat inference ➡️Iterate to see if you can speed up training!

🎙️ The fastest inference for realtime voice AI agents is now on Together AI Ultra-low latency STT, sub-second TTS, and serverless hosting for open-source TTS models — all on one platform.

> 200 – 300 sequential tool calls without human interference actually INSANE if true/translates to real world usage for long-horizon tasks

RAG is not so dead, after all it seems...

Yes, K2 thinking recommended temp = 1.0