Field report: 0.6.0 → 0.6.2 validation numbers from a 24/7 production fleet (M3 Ultra + M4 Max)

[anicaise-ai]

Some validation datapoints from our fleet after upgrading, in case they're useful:

v0.6.2 vs v0.6.1, Qwen3.6-35B-A3B-mxfp8 (MTP on, hybrid GDN), M4 Max 128 GB, A-B-A protocol (0.6.2 → 0.6.1 → 0.6.2, fresh unique long prompts per phase to defeat the persistent SSD prefix cache):

• MTP backbone time per cycle: 19.0 / 19.8 / 19.6 ms — equal within noise;
• MTP acceptance: 78.8 / 76.6 / 78.4%;
• cold prefill (~14–15k tokens): 1505 / 1528 / 1527 tok/s;
• temp-0 outputs byte-identical across versions (6/6 fingerprint prompts, including a 12k-token needle).

No regression anywhere; 0.6.2 is our production build on this host now.

Earlier confirmations of your announced gains, for the record:

• 0.6.0 on DeepSeek-V4-Flash (M3 Ultra): long-prefill reference 21.2 s → 19.7 s cold (SSD cache invalidated, so that's a floor); 0.6.1 brought it to 13.6 s.
• 0.6.1 on the Qwen 35B (M4 Max): decode 149 tok/s short and 146 tok/s at 39k cached tokens (+17% vs 0.6.0) — the MoE-router decode win is real at long context too.
• The 0.6.x reasoning_content separation eliminated a whole class of "reasoning fragments reaching users" bugs for us on complete responses (see separate issue for the finish=length edge case).