Modelpublic

philipjohnbasile/hy3-demolition-mlx-reap25-v1-mtp

sourceHugging Faceapache-2.0updated 16d agoView on Hugging Face

2likes228downloads

Model Card

Hy3-Demolition-MLX reap25-v1-mtp

Explore the model guide · All public work

Release at a glance

	This artifact
Purpose	The REAP25 trunk with a draft sidecar whose expert axis was reduced to match the 144-expert trunk.
Runtime	Requires a compatible Hy3 implementation and MTPLX MTP contract. Structural recognition is separate from generation qualification.
Status	Runtime verification pending; see the evidence and limits below.
Tensor download	86.89 GB (80.92 GiB) of root `.safetensors` files, including any root sidecars. This is a file-size total, not peak RAM.
Read first	Draft-head modification can change acceptance; output equivalence depends on a correctly implemented and verified decoding path.

The MTP-equipped variant of reap25: its fused trunk with the Hy3 NextN (Multi-Token-Prediction) sidecar grafted back on (num_nextn_predict_layers=1, num_experts=144), for self-speculative decoding on MTPLX.

Runtime status — reviewed September 10, 2026

The original recognized-backend-pending observation came from MTPLX 2.0.1. It is historical: the maintainer subsequently confirmed that the Hy3 and Qwen MTP backend work shipped in 2.1.0 through the release branch. Read the upstream shipping record.

That code shipment does not qualify this exact checkpoint. The source Hy3 implementation, runtime contract, loading, output agreement, and performance still need to be checked together. No fresh end-to-end MTP qualification of this artifact is claimed here. Upstream mlx-lm PR #1211 remained open at review.

The documented AR path is the sibling model.

How it was built

The MTP head consumes the trunk's final hidden state (hiddensize 4096, unchanged by pruning) and runs its own MoE on the global `numexperts. So the base checkpoint's mtp.*` sidecar grafts directly onto the fused AR trunk — no re-heal, no re-prune of the trunk.

—The MTP head's experts are sliced 192→144 (packing-safe whole-expert axis slice, scales-L2 pick) to match the pruned trunk's num_experts. The MTP head is a draft component MTPLX verifies every token against, so this changes the draft proposal distribution. Output equivalence requires a correctly implemented and validated verification path; it is not established by grafting the head.

Graft script + mtplx inspect receipts: https://github.com/PhilipJohnBasile/hy3-demolition-mlx (scripts/38_mtp_sidecar_graft.py, eval/receipts/mtplx_inspect_*.json).

Limitations

—Does not run on stock mlx_lm as MTP. The fast MTP path needs the MTPLX backend; mlx-lm's own per-token self-speculative loop is ~4.7× slower than AR (measured), which is why the MTPLX batched-verify backend is the target.
—Everything from the base reap25 card applies (quantized MoE, English/agent focus, no tool execution).
—End-to-end MTP behavior is unverified until the backend loads it; recognition is structural (mtplx inspect), not a live run.

Base recipe + receipts: https://github.com/PhilipJohnBasile/hy3-demolition-mlx

Other serving applications

No current LM Studio or Ollama qualification of this artifact is recorded here. An architecture becoming available in one library does not establish support in every application. Use the AR sibling's pinned release recipe as the documented starting point.

Experimental streaming

The source project's SSD pager measurements concern an AR serving path. They do not qualify this MTP variant on a smaller-memory machine. Read the AR streaming experiment.