# MTPLX Forge and MTP-preserving MLX conversion > 2.9.2 (25 Aug 2026): Forge correctness fix for the MTP norm convention. Parent: [Mac local LLMs: Speculative decoding and MTP](https://llms-explorer.com/tree/mac-local-llms-speculative-decoding-and-mtp/) · 1 facets · 20 facts · page: https://llms-explorer.com/tree/mtplx-forge-and-mtp-preserving-mlx-conversion/ ## Facts - 2.9.2 (25 Aug 2026): Forge correctness fix for the MTP norm convention. — source: `asserted` - 2.12.0 (Sep 2026): Forge converts Flash-Next source models (issue 508). — source: `asserted` - The home page describes the flow as "paste a Hugging Face link". — source: `asserted` - A wrong norm convention collapsed draft acceptance to 0-2% on some packs until 2.9.2. — source: `asserted` - A trunk shifted twice (norm offset applied twice) is now refused by the runtime with a clear error. — source: `asserted` - A Forge verdict can be that MTP does not help: the tool reports "Depth 1 is fastest" style results from measurement rather than assuming a gain. — source: `asserted` - None found. All Forge claims are first-party (MTPLX README and release notes); no third party describes using Forge. — source: `asserted` - What data and how many steps Forge uses to train the MTP adapter is not in the cached pages. — source: `asserted` - Whether Forge can preserve an existing in-checkpoint MTP head losslessly, as against training a new adapter, is not stated. — source: `asserted` - Third-party reproduction of Forge speedups is missing. — source: `asserted` - MTPLX Forge turns a Hugging Face repo into an MTPLX-ready MTP model by converting to MLX, training the MTP adapter, verifying the result is faster and still exact, and optionally publishing to the Hub. — [source](https://github.com/youssofal/MTPLX) - `mtplx forge --help` lists probe, build, publish and verify subcommands, and Forge is also a screen in the Mac app. — [source](https://github.com/youssofal/MTPLX) - Pulls and Forge builds always write to the primary model root; extra `model_dirs` are discovery-only, and CLI updates or removals in an extra root need `--cache-dir` naming that root. — [source](https://github.com/youssofal/MTPLX) - MTPLX 2.9.2 (25 Aug 2026) decides the MTP norm convention once per tensor set, which rescued packs whose draft acceptance had collapsed to 0-2%. — [source](https://mtplx.com/releases/) - Since 2.9.2 the MTPLX runtime refuses a double-shifted trunk with a clear error instead of running it. — [source](https://mtplx.com/releases/) - MTPLX 2.12.0 lets Forge convert Flash-Next source models (issue 508, contributed by bpmforge). — [source](https://mtplx.com/releases/) - The MTPLX home page describes Forge as "paste a Hugging Face link; Forge converts it to MLX and measures the speedup on your Mac." — [source](https://mtplx.com/) - The official Youssofal catalog includes FP16 builds of Qwen 3.8 27B for M1 and M2, a Qwen 3.6 35B MoE balance build, and Gemma 4 packs. — [source](https://github.com/youssofal/MTPLX) - A merge commit in the MTPLX repo lists `mtplx/commands/forge.py`, `compressed_tensors.py`, `gdn_capture.py` and `mtp_patch.py` as the conflicting modules of the 2.9.2 staging merge. — [source](https://github.com/youssofal/MTPLX) - The MTPLX repo's CI fails any push whose history contains AI authorship or co-author attribution. — [source](https://github.com/youssofal/MTPLX)