Zhipu launches GLM-5.3-FlashX at 200 tokens/s, with its inference infra built by LLM agents
赛博禅心 · wechat · 2026-09-18
Zhipu released GLM-5.3-FlashX, offering up to 200 tokens/s — 5x faster than GLM-5.3-Flash at 2.5x the price — running on 100,000 domestic chips, with the API now live.
The standout detail: the inference infrastructure was built largely autonomously by agents driven by GLM-5.3, going from model adaptation to production in under two weeks and tripling end-to-end throughput. Humans only set goals, boundaries, feedback environments, and reviewed critical changes; testing, logs, traces, and microbenchmarks were all delegated to agents in a dense-feedback loop.
Three concrete agent-driven fixes are documented: (1) FP32 inputs silently computed in TF32 in two tl.dot calls causing precision drift in long-context KDA paths, fixed with inputprecision="tf32x3" and merged upstream; (2) KVTransfer stalled by unreleased GIL in DeepEP v1.2.1 C++ calls, fixing a >20% performance gap down to under 1%; (3) redundant quadruple FP32 normalization/gating in KDA Decode, consolidated into one thread block for a 1.71x operator speedup.
More from coding & agent
- Dev builds 3D universe explorer with 20 galaxies, 767 planets using Codex GPT-6.1 Sol — OpenAIDevs · 2026-11-02
- ChatGPT, Claude and Perplexity's fetchers don't run JavaScript, leaving client-rendered sites invisible to AI — gaganghotra_ · 2026-11-02
- SGLang boosts multimodal inference >10% by replacing GPU memory bookkeeping with a dict lookup — charles_irl · 2026-11-02
- kivi: hold Fn, talk, and it types polished text anywhere on your Mac — itsOmSarraf_ · 2026-11-02
- A base LLM plus 11 lines of code is all you need for a decision model — helloiamleonie · 2026-11-02
- Fix codebase structure before asking AI agents to write tests, says engineer — arpit_bhayani · 2026-11-02