mohammadhaishemkhawaja/hashyy is a llama.cpp expert streaming project that states it processes an unpruned 177B MoE model at 11.5 tok/s on a single 12 GB GPU. Its description says it also carries the measurement harness and the failures it went through along the way. The 2026-10-03 commit 6189bf3 is a documentation change that touched only the one file README.md rather than any code, and consists of nothing but 17 added lines and 1 deleted line. What was changed is a single row of the measurement table, and the commit states that it is a change explaining what that row measures.
The README replaces the existing wording 'as inherited' with 'baseline' and insists that the baseline is not stock llama.cpp. It explains that this project already contained two rounds of work before the change the README covers, and lists them as the PR #25294 fork, the rewritten Windows read path, the stored heat map, and --poll 0. The measurement conditions are written down as well: on the same binary, restoring the old single shared file handle with LLAMA_MOE_STREAM_ONE_HANDLE=1, 9 I/O threads, without --poll 0, without a hot file, and with -c 4096, the before and after two arms were run alternately. It notes that the results fluctuate 3-5% from the OS page cache alone, and concludes that comparing with numbers obtained from other builds or on other days means nothing in particular.
In this document, 8.14 tok/s is a value that was previously cited for this project, but the README clarifies that it was the warm measurement under the looser earlier protocol, and that the 5.85 – 7.07 in the table are cold measurements taken with bench/sweep.py --cold. It states explicitly that only cold-to-cold is the comparison cited in this file. The remaining values in the table are 11.43 – 11.58 when alternating between different prompts, 11.55 – 11.57 for a cold start under the same conditions, and 9.1 → 13.4 (median 10.30) when a new conversation is carried forward turn by turn, and the summary below the table is ~7.0 → 11.6 tok/s, +65%.
That said, this change is explanatory only. There is no performance improvement and no code fix, so it cannot be read as meaning the repository has become faster. The numbers in the table and the +65% summary are all values the repository itself compiled, not independent verification. Under the record that the results fluctuate 3-5% from the page cache alone, it should be kept in mind together that placing the warm 8.14 tok/s and the cold 5.85 – 7.07 side by side means the conditions differ. Rather than carrying the numbers over as they are, it is better to first read which conditions each value came from.