github beamivalice/sushi v1.0.0
sushi v1.0.0

latest releases: v1.6.1-dev1, v1.6.0, v1.6.0-dev3...
one month ago

v1.0.0 — Qwen3.8-Flash-Next, sushi-packed

Add support for 2 sushi models

Quality

KLD vs size

KLD against the bf16 model: 16 prompts x 512 tokens scored to the first EOS, kv8, every pack run by the same sushi
build. The light rings are the sushi packs with a 4-bit n-gram table (Sushi-3bpw ships that table; either table works
with either pack). Numbers: docs/quality-kld.md.

Speed

Sushi-3bpw on an M5 Max 128 GB, sushi v1.0.0 release candidate (build 725b76c): --ctx-size 1048576 --kv-quant 8 --mtp, llmprobe --bench-only, quiet box.

decode and prefill vs context

Don't miss a new sushi release

NewReleases is sending notifications on new releases.