Two Flash-Next and GLM-5.3 Flash oMLX community sessions were uploaded on 12 October GMT and checked against their original result pages today. Both report an M5 Ultra with 64 GPU cores and 256GB. These are separate model, quantization and engine recipes, not a controlled cross-model race.
The original GLM-5.3-Flash-oQ4e-mtp session uses oMLX 0.7.1.dev1 and Lightning MTP depth 3. Its seven buckets span 1K 1,261 PP / 68.7 TG to 195K 1,807 / 65.4 tok/s. The 16K row reports 1,917 PP / 138.8 TG, much higher than adjacent 8K (75.9) and 32K (94.3). It is a single source result; its cause and repeatability are unverified, so we do not turn that spike into a stable baseline or payback assumption. The separate four-stream total is 121.3 TG tok/s, with per-stream contexts unpublished.
Prices were rechecked today: Flash remains US$0.15 / $0.47 per million input/output tokens; the calculator's 27B Near AI US$0.04 / $1.35, 35B US$0.05 / $0.70, and GLM Relace US$0.04 / $0.50 pairs remain listed. They are editable market proxies, not booked revenue. The Flash-Next license requires separate permission for commercial Model as a Service.
今天重新核對市場牌價：Flash 每百萬輸入／輸出 token 仍是 US$0.15／$0.47；試算中的 27B Near AI US$0.04／$1.35、35B US$0.05／$0.70，以及 GLM Relace US$0.04／$0.50 仍在頁面上。這些是可調整的市場參考價，不是已實現收入。Flash-Next 模型授權要求商業 Model as a Service 另外取得許可。