SYLIN

Needle

by Cactus Compute · Model · Language & code models

8-29 MB on-device tool-calling model, 29M-121M parameters, with 2-20 selectable layers; beats models 10x its size

AccessOpen weightsDownload0 GBSize0.045B parametersLicenseMIT

Versions

VersionReleasedRank
Needle 3LATESTSep 20
Needle 2Aug 16

Evidence

  1. Sep 20
    Claimed to beat models 10x larger on mobile tool-calling benchmarks
  2. Aug 16
    Needle 2: 45M-parameter tool-calling model, 14 MB binary, 28 MB RAM; 500 tok/s on Raspberry Pi 5