GLM-5.3-Flash-Alis-MTP-Drafter

GLM-5.3-Flash-Alis-MTP-Drafter (bf16)

The standalone native MTP (nextn) drafter of GLM-5.3-Flash, extracted from the official FP8 release (84c6a6aa) for self-speculative decoding with mlx-vlm PR #2044. To our knowledge the first standalone GLM-5.3-Flash MTP drafter on the Hub — normal MLX conversions drop layer 45, so quantized targets otherwise have no MTP path.

• 14 GB, 29 tensors (mtp.*), bf16 (+ fp32 router correction bias), model_type: glm5_next_mtp, block size 2.
• Loads as Glm5NextMTPDraftModel via load_drafter(path, kind="mtp") on the PR #2044/#2074 branches.
• Measured (M3 Ultra 512GB, 4-bit-class target, ctx 512, greedy): 1.12× decode, bit-identical to plain decode, ≈89% single-draft acceptance.

Why a dequantized extraction

The PR branch's splitter currently writes raw FP8 + unstacked weight_scale_inv tensors from the FP8 source and drops the quantization config, producing an unloadable artifact (reported upstream: #2044 comment). This repo applies the reference blockwise dequant (block 128, mx.from_fp8, per-expert scale grids rejoined across the stacked switch_mlp) to publish a clean bf16 drafter. Receipts in-repo: SPLIT_RECEIPT.json (extraction) and DEQUANT_RECEIPT.json (post-process).

from mlx_vlm.speculative.drafters import load_drafter
drafter, kind = load_drafter("avlp12/GLM-5.3-Flash-Alis-MTP-Drafter", kind="mtp")

MIT, derived from the MIT base release.