Unsloth released version 0.1.921-beta on 2026-11-20. The build makes image generation faster on 8 GB GPUs, dropping a MiniMax‑H3‑Image prompt from 82.7 s to 36.7 s. GGUF image models now need 52‑66 % less RAM, and FLUX.1 steps run 16‑28 % quicker when streamed from system RAM. Macs with Apple Silicon can restyle whole video clips, and MCP‑enabled chats now only reprocess new tokens, reducing work by up to 87 %.
Why it matters
People with modest GPUs can create images and run chats noticeably faster, saving time and memory.