Bonsai 2 27B runs in 5.9 GB, under Apache 2.0

TL;DR

  • On 17 September PrismML released Ternary Bonsai 2 27B, a version of Qwen3.8 27B whose weights are -1, 0 or +1.
  • It takes 5.9 GB instead of 53.8 GB and scores 83.9 against the original's 85.4 on PrismML's 20-benchmark suite.
  • It needs PrismML's own llama.cpp fork or its MLX package to run.

Read the full story

Sign in with your email to read AI News. It’s free.

Share this story

Explain like I’m 15