---
title: MiniCPM-V-4.7-35B-A3B-gguf-hexagon
canonical_url: "https://www.modelscope.cn/models/fengerhu1/MiniCPM-V-4.7-35B-A3B-gguf-hexagon"
md_url: "https://www.modelscope.cn/models/fengerhu1/MiniCPM-V-4.7-35B-A3B-gguf-hexagon.md"
repository: fengerhu1/MiniCPM-V-4.7-35B-A3B-gguf-hexagon
last_updated: 2026-10-15
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
library_name:
  - gguf
downloads: 195
stars: 0
tags:
  - gguf
  - hexagon
  - minicpm-v
---

# MiniCPM-V-4.7-35B-A3B-gguf-hexagon

> MiniCPM-V-4.7-35B-A3B-gguf-hexagon - fengerhu1 在 ModelScope 开源的模型。MiniCPM-V 4.7 mixed-Q3 LES weights, Hexagon tiled experts, Q8 DSpark and F16 vision projector

fengerhu1/MiniCPM-V-4.7-35B-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/MiniCPM-V-4.7-35B-A3B-gguf-hexagon
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Tags**: gguf, hexagon, minicpm-v
- **Downloads**: 195
- **Stars**: 0
- **Last updated**: 2026-10-15

Source: https://www.modelscope.cn/models/fengerhu1/MiniCPM-V-4.7-35B-A3B-gguf-hexagon

---

# MiniCPM-V 4.7 35B-A3B GGUF Hexagon

Derived from https://modelscope.cn/models/OpenBMB/MiniCPM-V-4.7-35B-A3B-gguf . Matching runtime: https://github.com/yydhYYDH/llama.cpp-omni/tree/hexagon-minicpm-dspark .

This is the deployed OnePlus 13 mixed-Q3 LES configuration, not Q4_K_M. Download all files together. Main GGUF contains non-expert weights and expert metadata; expert payload is separate and required.

| Asset | Description |
| --- | --- |
| minicpm47-les-hexagon.gguf | Hexagon-prepared non-expert GGUF |
| experts-htp-tiled/layer_000.bin through layer_039.bin | 40 layers of Q3_K experts, Hexagon tiled backend layout |
| experts-htp-tiled/manifest.json | Expert metadata; original source paths are provenance |
| dspark-MiniCPM-V-4_7-q8_0.gguf | Q8_0 DSpark draft quantized from original F16 |
| mmproj-MiniCPM-V-4_7-real-f16-pad4320.gguf | Original projector, BF16 converted to F16, FFN padded4304 to4320; not 4x |
| SHA256SUMS | File integrity checks |

Prepacked experts require this matching backend; they are not standalone GGUF or portable CUDA weights. Set LLAMA_MOE_HTP_PACKED_DIR to experts-htp-tiled and pass --moe-experts experts-htp-tiled. Both target and draft use HTP0,HTP1. Build Android executables and libraries separately.

Processing includes expert extraction/packing, draft quantization, and vision conversion/padding; no retraining. Short text and image cases have been validated, broad accuracy equivalence is not established. Upstream Apache-2.0 license applies.