🌔 Work at Moonshot.ai (Kimi)
Post-training on K3 Vision Agentic · K2.6 VTIR · K2.5 Vision Perception & Knowledge & Reasoning
2026.07
Kimi K3
Open-weight flagship · Frontier Vision Agentic · 2.8T parameters · native multimodal · 1M context
Vision Singlestep & Agentic RL
- Long-run RL for vision singlestep and agentic, for multiple rounds of iterative optimization (K3 Report Chap4.2.3)
- Optimization on multiple harnesses generalization
- Scaling Mid-train data and crafting RL Promptset
- At launch: Top-2 on Vision Agentic, e.g.: CharXiv RQ w/ python, only behind Claude Fable 5. Generalized vision agentic capability on multiple harnesses.
2026.01 / 04
Kimi K2.5 / K2.6
Kimi's First Unified Vision–Text Model -> First Opensourced Frontier-level VTIR
Vision Post-training & VTIR
- RL on Zero-Vision SFT: Text-only SFT model, followed by a large and long-run RL, activates close-to-SOTA vision capabilities. And then producing lots of on-policy vision data. (K2.5 Report Chap2.2, ZeroVision ColdStart -> Vision-Centric RL)
- Coldstart & RL on K2.6 VTIR: enhancing visual perception, understanding and reasoning capabilities by general ipython integration, achieving Top-3 VTIR performance (Top-1 in Opensourced).
- Large-scale Chart-to-Code
