Unsloth turns on multi-token prediction for two open models

Unsloth turned on multi-token prediction by default for two open MoE models, Qwen3.8-Flash-Next and GLM-5.3-Flash, in the 0.1.805-beta release. The library claims up to 2x faster generation as a result.

For anyone running open models locally, generation speed decides whether the setup is usable day to day. A default speedup this size changes which models make sense to run on a laptop rather than in the cloud. So far, only Unsloth's own benchmark backs that 2x claim.

The same release adds fine-tuning for both MoE models on text or image datasets, through Apple's MLX framework. That lets Mac users train Qwen3.8-Flash-Next and GLM-5.3-Flash locally.

Follow-up fixes brought more changes on top of the beta. Long Qwen chats on Mac run up to 30x faster, and MLX models use their full context size instead of a truncated one. GLM-5.3 MLX fine-tunes export to GGUF for use outside MLX. Unsloth also added support for Codex's apply_patch tool, letting locally run models make edits the same way Codex does.