长 agent 会话接着聊，以前经常要从头 prefill。两台 DGX Spark 上的 GLM-5.3-Flash 配方把缓存修好了，续上会话的时间差得很开：
This is a AI post classified by AGK Intelligence as AI infra & evals (a launch), kept by the AI Radar because it carries real work, not commentary.
长 agent 会话接着聊，以前经常要从头 prefill。两台 DGX Spark 上的 GLM-5.3-Flash 配方把缓存修好了，续上会话的时间差得很开： 1. 450k token 的对话，重新 prefill 要 292 秒 2. 打开 CACHE_FIRST 之后，命中率 99%，续上下一轮只要 5.1 秒 3. Mia 那边测的是 380k 对话，从 4 分钟压到 3 秒 工具调用和 JSON 返回也修了一批。 我两台 Spark 跑的就是 GLM-5.3-Flash，长会话这块准备去更新。
Posted by Wei (7.1k followers) 16 h ago · 10 likes · 949 views · view the original post on X. Kept by the AI Radar as AI infra & evals.
More AI work like this
- The enterprise AI bottleneck has moved from models and GPUs to the data path. — @Dell
- AMD Localmaxxers are getting serious about inferencing. 🤯 — @TeksEdge
- We evaluated Nexus on 7 benchmarks and 1,222 tasks. It defines the Pareto frontier. — @FireworksAI_HQ
- There is a 560B model running with its full 1M context on 2x DGX Sparks by @andersjw_. — @plotarmordev
- 免费GPT的API，手快有手慢无 — @mylifcc
- TensorFold 1.2.4 is out. 🚀 — @ashxhart
- AWS is making it easier to run several smaller models on the same GPU. — @markfenner
- The @nvidia Jetson Orin Nano is back in the lab. — @jmurillocode
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 65.2k posts from 5k X accounts over the last 14 days, 10.6k tools, 19 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-07 16:45 UTC. Full method.