New Ultra context results.
The original oMLX Qwen3.8-Flash-Next session was uploaded on 29 September GMT. It used a 64-GPU-core M5 Ultra with 256GB, Q4e quantization, oMLX 0.7.0rc1, adaptive MTP depth 3 and thinking enabled. The configured 64K row reports 3,170 prefill and 102.4 single-stream generation tokens/s.
Context changes the answer
In that same Flash-Next session, 32K shows 3,216 PP / 67.6 TG, and 128K shows 3,101 / 102.1. These are configured context sizes, not published exact prompt token counts. Actual generation starting context, output length and per-run cache use were not disclosed. DFlash and ANE prefill were off. The 32K generation dip stays visible rather than being smoothed away.
A separate Swift-Qwen3.8-27B oQ4e-fp16-mtp session, uploaded 29 September 20:07 GMT (30 September in Taiwan), reports 1,064 PP / 43.4 TG at configured 1K and 1,860 / 49.0 at 32K. Its configured 1K eight-stream batch reaches 143.5 TG tok/s for the whole device; that is not per-stream speed. The site's PP/TG chart pairs only rows from the same run. The session's TTFT is milliseconds, not a count of output tokens; output lengths were not published. ANE prefill was explicitly off.
Three open code changes, three distinct workloads
Author tests for context-copy MTP PR #4104 show a copy-heavy file edit going from 190.0 to 247.9 single-stream TG, while four generic coding prompts were essentially flat (178.3 to 177.9). The 16-row verify PR #4105 reports 286.5–291.2 on the same kind of file edit in an integration build that also includes other PRs; the chart uses the lower 286.5 sample. Batch PR #4106 reports 257.3 to 293.3 TG for eight streams combined, with MTP off and short coding prompts. None is a general single-stream speed claim, and these open PR results are not paired with a separate prefill run.
Explore both charts