The MTP-equipped variant of hy3-family-mini-qwen35b-v1: our verifier-healed Qwen3.6-35B-A3B trunk with the Qwen NextN / MTP head (mtp.*, from the clean mlx-community/Qwen3.6-35B-A3B-MTP-4bit) grafted on for self-speculative decoding.

Status — MTP-recognized, runtime load pending (honest)

mtplx forge probe recognizes this artifact (has_mtp_weights: true, forgeable: true, backend qwen3_next, ~25 GB peak). However, mtplx forge build currently fails to load it — MTPLX's bundled mlx_lm raises Model type qwen3_5_mtp not supported. So, like the Hy3 MTP variants, this is staged, not yet verified-runnable: the MTP weights and metadata are correct and recognized, but the runtime load path for qwen3_5_mtp isn't there yet in the release. Use the AR sibling for actual inference today.

bash

mtplx forge build --repo <this-repo> --recipe '{"body_bits":4,"body_group_size":64,"body_mode":"affine","mtp_policy":"keep_bf16"}' ...
mtplx tune --depths 1,2,3     # find the best speculative depth for your Mac

What it is

Trunk: mlx-community/Qwen3.6-35B-A3B-4bit + our verifier-filtered LoRA heal (val 0.639, 8/10 executed stress — see the AR sibling card).
MTP head: the clean Qwen NextN head (a draft component MTPLX verifies every token against, so it never affects correctness — only draft acceptance/speed).
Clean base — no private domain-specific adapters are fused into this build.

Notes

The fast AR daily-driver path is the AR sibling; this variant adds MTP speculative decoding on MTPLX.
Graft script + receipts: https://github.com/PhilipJohnBasile/hy3-demolition-mlx

philipjohnbasile/hy3-family-mini-qwen35b-mtp-v1

Hy3-Family Mini — Qwen35B MTP v1

Status — MTP-recognized, runtime load pending (honest)

What it is

Notes

philipjohnbasile/
hy3-family-mini-qwen35b-mtp-v1