Mini‑GGUFs for Qwen3.6 Models Reduce Size from 38GB/29GB to 900MB/450MB

A new script extracts only the tensors needed for MTP conversion from large GGUF model files, producing two smaller GGUFs that are fully compatible with the original conversion script. The 900MB file (Qwen3.6‑35A3B‑MTP‑TENSORS‑ONLY) replaces the 38GB full model, while the 450MB file (Qwen3.6‑27b‑MTP‑TENSORS‑ONLY) replaces the 29GB full model. Both mini‑GGUFs were verified against SHA256 hashes and produce identical results to the full models during conversion tests. The reduction in size dramatically speeds up downloads and saves bandwidth for users who only need the tensors for conversion. The authors note that the MTP implementation is not finalized, so the models may become obsolete or break with future updates. They advise keeping the original full models in case the conversion process changes. The project credits the original GGUFs from am17an and the conversion script from buzz. The mini‑GGUFs are available on Hugging Face under the Qwen3.6‑35A3B‑MTP‑TENSORS‑ONLY and Qwen3.6‑27b‑MTP‑TENSORS‑ONLY repositories.

This release is a practical tool for developers working with large language models who want to reduce storage and bandwidth requirements without sacrificing conversion compatibility. It demonstrates a lightweight approach to model manipulation that could be adapted to other model families and conversion workflows.

The article is a technical tip rather than a major product launch, but it offers a concrete improvement for AI developers and hosting teams looking to optimize model handling.