---

library_name: llama.cpp

license: other

license_name: qwen-community-1.0

license_link: https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE

base_model: Qwen/Qwen3.8-Flash-Next

base_model_relation: quantized

tags:

- gguf

- mtp

- nextn

- speculative-decoding

- draft-model

- rocmfp4

- rocmfpx

- strix-halo

- qwen4exp

pipeline_tag: text-generation

quantized_by: pugant

version: '1.0'

date: 2026-09-10

---

Qwen3.8-Flash-Next MTP draft head — head-only ROCmFP4 requant (GGUF)

The jointly-trained MTP (NextN) draft head of Qwen/Qwen3.8-Flash-Next (arch qwen4exp), requantized head-only for ROCmFPX engines: the draft LM head (output.weight) is converted to Q4_0_ROCMFP4 (644.14 → 341.02 MiB), while all other 33 tensors are byte-identical to the community Q8_0 extraction of the official head.

This is the sidecar served in our production engine since 2026-09-10.

> About the filename: the file is named …-MTP-Q8_0-HEAD-ROCMFP4.gguf because the base

> file type is Q8_0 — only the head tensor is ROCmFP4. The per-tensor types are the

> authoritative ones (same convention as our model cards).

Why a matched-quant draft head

The draft head re-reads its entire LM head (248k vocab × 2560) for **every drafted

token** — a per-token matmul over the full vocabulary. On a bandwidth-limited APU that

read dominates the draft traffic. Quantizing the drafter like the target has a

measurable side benefit beyond the byte saving: the two quantized models are wrong in the

same places, so acceptance goes up, not down (first observed by

agentionai

on the same hardware class).

Measured

Radeon 8060S (gfx1151), Vulkan, target = our

ROCmFP4 STRIX_LEAN

build, draft-mtp n-max 6 / p-min 0.75, interleaved A/B in a production-style container.

Full protocol in the lab wave-6 note.

| sidecar | decode t/s | acceptance |

|---|---:|---:|

| Q8_0 head (community) | 23.92 | 0.652 |

| this head | 24.99 (+4.5%) | 0.667 |

Greedy fidelity: the engine release this head ships in passed the 12-prompt battery

bit-identical to the pinned canon (sha 6cc4f583…) — the head-only requant does not

change generated text in the greedy regime.

Runtime

The Q4_0_ROCMFP4 tensor type needs a ROCmFPX-capable build — stock llama.cpp cannot

load this file. Recommended runtime: our lab build

(pugant/strix-nebulosa, main) — full source,

Dockerfile and guides; upstream format definition:

charlie12345/ROCmFPX.

llama-server -m Qwen3.8-Flash-Next-ROCmFP4-STRIX_LEAN-*.gguf \
  -md Qwen3.8-Flash-Next-MTP-Q8_0-HEAD-ROCMFP4.gguf \
  --spec-type draft-mtp --spec-draft-ngl all \
  --spec-draft-n-max 6 --spec-draft-p-min 0.75

Pairs naturally with the LEAN FP4 target above (same quant family on both sides of the

speculative pair — that is the point of the matched quant).

Provenance & reproduction

• Base: the community Q8_0 extraction of the official head

(quimmedes/Qwen3.8-Flash-Next-MTP-GGUF),

itself converted from the official checkpoint via the llama.cpp NextN export

(PR #27836 lineage).

• One tensor converted (output.weight Q8_0 → Q4_0_ROCMFP4); the 33 non-head tensors

are sha256-identical to the base; all metadata values unchanged (file size 3.82 GB).

• Reproduce (native quantizer flags, no patches needed — quantize flags go first):

```bash

llama-quantize --allow-requantize --pure \

--output-tensor-type Q4_0_ROCMFP4 \

mtp-Qwen3.8-Flash-Next-Q8_0.gguf \

Qwen3.8-Flash-Next-MTP-Q8_0-HEAD-ROCMFP4.gguf Q8_0

```

Script with dry-run pre-flight and stat checks:

scripts/w6t3-drafter-head-requant.sh

in the lab snapshot; it regenerates this artifact bit-identically.

• sha256: fa30a89fa06cd1bcc40810dcf1fed73d69c5ff33afb2a08b76333c6f9629ea34

Credits & license

• Qwen team — the MTP head itself, trained jointly with the target model.
• quimmedes — the Q8_0 GGUF extraction this file

derives from, and the cafe-llama.cpp fork that runs it.

• charlie12345 (ROCmFP4 format / ROCmFPX) and

ciru-ai (ROCmFPX) — the tensor types.

• danielhanchen (qwen4exp support,

PR #27742) and JJJYmmm

(MTP draft graph, PR #27739).

• agentionai — the matched-quant observation.

Derivative of Qwen/Qwen3.8-Flash-Next, distributed under the **Qwen Community License

1.0** (see LICENSE; note the license's Model-as-a-Service clause before commercial

serving). Requantization and measurements by pugant; work assisted by GLM by z.ai.

---

> Everything here is experimental and provided as-is, at your own risk.