Swift Bonsai 2 MTP GGUF
The base is Swift Bonsai 2, UkisAI's reasoning-efficient derivative of Prism ML's Ternary Bonsai 2 27B, which they report uses ~40% fewer thinking tokens on GPQA-Diamond while holding accuracy roughly even. This release grafts on ProCreations' MTP head for speculative decoding, so you get both: the shorter reasoning traces of the Swift base and the drafter from the ProCreations head.
This is a plain byte-merge: the MTP head from ProCreations/Ternary-Bonsai-2-27B-MTP grafted onto the base from ukisai/Swift-Bonsai-2-GGUF. No other work has been done on this file.
Change from upstream: the ProCreations MTP head was grafted onto the ukisai Swift base. No other weights, metadata, or behavior were altered.
License and attribution
This repository is not an official Prism ML, Qwen, or DFlash release. The base releases (ukisai and ProCreations) are community derivative releases of the official Prism ML Bonsai and Qwen models, not official releases or endorsements.
This model is distributed under Apache-2.0. The NOTICE.txt retains the upstream attributions for Prism ML (Bonsai), Qwen/Alibaba Cloud, and ProCreations (MTP head adaptation), plus a change statement for this merge. As of the date of writing this README (2026-10-25), UkisAI had added nothing to its NOTICE.txt — this is a point-in-time snapshot and may become outdated if that upstream file is later updated. The NOTICE.txt in this repository is the authoritative copy of the attributions that ship with this release.