---

base_model: zai-org/GLM-5.3-Flash

base_model_relation: quantized

license: mit

library_name: gguf

tags:

- gguf

- glm5-next

- mtp

- speculative-decoding

---

GLM-5.3-Flash MTP draft head (GGUF, Q4_K)

The GLM-5.3-Flash multi-token-prediction (MTP / NextN) block as a small standalone GGUF

(4.3 GiB), for speculative decoding next to a GLM-5.3-Flash model you already have. No need

to download a whole model again just to get MTP.

llama-server -m GLM-5.3-Flash-GSQ-RCO-3.0bit-q4kattn.gguf \
    -md GLM-5.3-Flash-MTP-Q4_K.gguf --spec-type draft-mtp

The draft borrows the token embeddings and output head of the main model, so only the MTP

block itself is stored here. The main model verifies every drafted token: the draft's

quantization changes how often drafts are accepted, never the output.

Requirements

• **The neurall/llama.cpp fork, release

b11327 or newer.** Stock llama.cpp does not run GLM-5.3-Flash MTP

yet (PR #27917), and loading an MTP-only

file as -md is a fork addition.

• A main GLM-5.3-Flash GGUF with the glm5-next architecture name, e.g.

pfeifferj/GLM-5.3-Flash-GSQ-RCO-GGUF

(3.0-bit, 3.5-bit) or

neuralll/GLM-5.3-Flash-GSQ-RCO-3.0bit-Q4Kattn-GGUF.

unsloth's GLM-5.3-Flash GGUFs use a different architecture name (glm5next) and don't load

in the fork.

Will it be faster?

It depends on how much of the model the GPUs hold. Every drafted token has to be verified,

and each token picks its own experts; experts that run on the CPU cost about as much per

drafted token as per generated one.

Measured on 2x RTX 3090 (48 GB) + 125 GB DDR4 with the fork's expert cache, 1500-token chat

reply, GLM-5.3-Flash 3.0-bit Q4_K attention (106 GB, 2.3x the VRAM):

| | decode t/s |

|---|---|

| without MTP | 20.2 |

| with this MTP head (best draft depth, 1) | 17.6 to 18.4 |

Draft acceptance is ~70-75% at depth 1, but here ~30% of expert work runs on the CPU and the

draft's VRAM shrinks the expert cache, so MTP is slower on this setup. Expect a gain when

most of the model fits in VRAM (more or bigger GPUs, smaller quants). For comparison, on the

same box Qwen3.8-Flash-Next (1.9x the VRAM, 95% of experts cached) gets 1.21x from MTP.

How it was made

The NextN block (blk.45.*, 29 tensors) is taken byte for byte from

unsloth/GLM-5.3-Flash-GGUF UD-Q4_K_XL

(attention Q8_0, experts Q4_K/Q5_K), fetched with HTTP range requests instead of downloading

the model. Metadata comes from the GSQ-RCO main model, with block_count 46,

nextn_predict_layers 1 and the MTP layer's per-layer values from unsloth's file. Scripts in

this repo (also in the fork under tools/moe-bench/, run from a fork checkout): gguf_remote.py (read a remote GGUF header) and glm_splice_mtp.py

(--mtp-only writes this file; without it, it writes a full model with the MTP block).

Credits and license

• Model and MTP weights: zai-org/GLM-5.3-Flash, MIT.
• Quantization of the MTP block: unsloth/GLM-5.3-Flash-GGUF, MIT.
• GLM-5.3-Flash MTP support in llama.cpp: PR #27917 (timkhronos).

MIT license, see LICENSE.