Tools
Unsloth 0.1.808-beta moves to PyTorch 2.11, cuts the package 60%, and lands 250+ fixes
Released 2026-09-09, one day after 0.1.807-beta, this build claims 1.2-1.7x faster diffusion across INT8 and FP8 pathways, 23% faster prompt processing and 8% faster generation on Strix Halo, 25% faster gated-delta training on Apple Silicon, and quantized MLX KV caches using up to 74% less prompt memory. PyTorch moves from 2.10 to 2.11 as the installer default, Windows llama.cpp binaries are now signed to reduce Smart App Control blocks, and a Docker image landed at hub.docker.com/r/unsloth/unsloth with native AMD64 and ARM64 variants.
Source
↳ Follow the thread