Chinese long context models

Models that keep accuracy at 128K+ tokens — Kimi, GLM-Long, Qwen2.5-1M. Useful for long-document Q&A, whole-repo code review, or multi-hour transcripts where chunking would lose the through-line.