Skip to content
MindPattern
Wire
Briefings
Search
Subscribe
← The Wire
Source trail
piszczek.pl / Hacker News
Public MindPattern findings, entities, and graph evidence that cite this source.
Findings
1
All-time hits
1
High value
0
Last seen
2026-08-18
Related findings
2026-08-18 / HACKER NEWS
Qwen3.8 27B Hits 50.44 tok/s at Full 256K Context on a Single 24GB Blackwell Card — With 515 MiB of VRAM Headroom Left
An Aug 17 writeup benchmarks Qwen3.8 27B on an RTX PRO 4000 Blackwell SFF using a custom 5.01 BPW iMatrix/NVFP4 hybrid quant (bulk matrices NVFP4, sensitive tensors Q5_K/Q6_K) and a patched llama.cpp with Gated DeltaNet support and chained MTP speculative decoding. Production mean was 50.44 tok/s versus 21.19 tok/s target-only greedy — a 2.81x speculative speedup — validated at 261,500 input tokens, occupying 23,952 of 24,467 MiB. The honest caveats matter: throughput collapses to 12.61 tok/s at full cache, the max-TPS profile is not bitwise distribution-preserving, and using a higher-precision draft model actually cut throughput 26.6%.
Open latest cited source
Wire
Briefings
Search
Subscribe