---
title: Qwen3.5-35B-A3B-gguf-hexagon
canonical_url: "https://www.modelscope.cn/models/fengerhu1/Qwen3.5-35B-A3B-gguf-hexagon"
md_url: "https://www.modelscope.cn/models/fengerhu1/Qwen3.5-35B-A3B-gguf-hexagon.md"
repository: fengerhu1/Qwen3.5-35B-A3B-gguf-hexagon
last_updated: 2026-10-17
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
library_name:
  - gguf
downloads: 14
stars: 0
tags:
  - gguf
  - hexagon
  - qwen3.5
  - dflash
  - speculative-decoding
---

# Qwen3.5-35B-A3B-gguf-hexagon

> Qwen3.5-35B-A3B-gguf-hexagon - fengerhu1 在 ModelScope 开源的模型。Qwen3.5-35B-A3B GGUF Hexagon

fengerhu1/Qwen3.5-35B-A3B-gguf-hexagon 是 ModelScope 魔搭社区上的image-text-to-text模型，采用 apache-2.0 许可。

- **Repository**: fengerhu1/Qwen3.5-35B-A3B-gguf-hexagon
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Tags**: gguf, hexagon, qwen3.5, dflash, speculative-decoding
- **Downloads**: 14
- **Stars**: 0
- **Last updated**: 2026-10-17

Source: https://www.modelscope.cn/models/fengerhu1/Qwen3.5-35B-A3B-gguf-hexagon

---

# Qwen3.5-35B-A3B GGUF Hexagon

Derived from Qwen/Qwen3.5-35B-A3B (Q4_K_M; Qwen3.5-35B-A3B is a multimodal model).
DFlash draft from z-lab/Qwen3.5-35B-A3B-DFlash. Multimodal projector from
ggml-org/Qwen3.5-35B-A3B-GGUF (Q8_0). Matching runtime: llama.cpp-omni Hexagon branch.

Main GGUF contains non-expert weights and expert metadata; the expert payload is separate
and required. The experts are Hexagon tiled weights (hexagon-tiled-v1 backend layout),
not standalone GGUF or portable CUDA weights.

| Asset | Description |
| --- | --- |
| models/qwen35-les-hexagon.gguf | non-expert text GGUF (experts stripped, 2.38 GiB) |
| models/mmproj-Qwen3.5-35B-A3B-Q8_0.gguf | multimodal projector (vision), Q8_0, 614 MB |
| experts-htp-tiled/layer_000.bin .. layer_039.bin | 40 layers of Q4_K/Q4_K/Q5_K experts, Hexagon tiled layout |
| experts-htp-tiled/manifest.json | expert metadata (backend_layout = hexagon-tiled-v1) |
| experts-htp-tiled/conversion.json | packing provenance |
| models/Qwen3.5-35B-A3B-DFlash-Q8_0.gguf | DFlash draft, Q8_0 |
| SHA256SUMS | file integrity |
| run-qwen35-text-fast.sh | example text run script (adapt paths) |

Prepacked experts require the matching Hexagon backend. Set LLAMA_MOE_HTP_PACKED_DIR and
--moe-experts both to experts-htp-tiled. Pass --mmproj models/mmproj-Qwen3.5-35B-A3B-Q8_0.gguf
for image input. Processing includes expert extraction/packing, draft GGUF conversion and
projector quantization; no retraining. Q4_K_M expert quant; target features used by the draft
come from the quantized target. Broad accuracy equivalence is not established. Upstream
Apache-2.0 licenses apply.