---
title: Sage-A1.5-gguf-hexagon
canonical_url: "https://www.modelscope.cn/models/fengerhu1/Sage-A1.5-gguf-hexagon"
md_url: "https://www.modelscope.cn/models/fengerhu1/Sage-A1.5-gguf-hexagon.md"
repository: fengerhu1/Sage-A1.5-gguf-hexagon
last_updated: 2026-10-21
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
library_name:
  - gguf
downloads: 25
stars: 0
tags:
  - gguf
  - hexagon
  - qwen3.5
  - moe
---

# Sage-A1.5-gguf-hexagon

> Sage-A1.5-gguf-hexagon - fengerhu1 在 ModelScope 开源的模型。Sage-A1.5 GGUF Hexagon

fengerhu1/Sage-A1.5-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/Sage-A1.5-gguf-hexagon
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Tags**: gguf, hexagon, qwen3.5, moe
- **Downloads**: 25
- **Stars**: 0
- **Last updated**: 2026-10-21

Source: https://www.modelscope.cn/models/fengerhu1/Sage-A1.5-gguf-hexagon

---

# Sage-A1.5 GGUF Hexagon

Derived from `InternScience/Sage-A1.5-Q4_K_M-GGUF` (Sage-A1.5 is a 35B-A3B
`qwen3_5_moe` agentic VLM, 40 layers, 256 experts, top-8). This bundle is
repacked for the llama.cpp-omni Hexagon expert-offload runtime.

The 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`), not standalone GGUF or portable CUDA weights.

| Asset | Description |
| --- | --- |
| models/sage-a1.5-les-hexagon.gguf | non-expert text GGUF (experts stripped) |
| models/Sage-A1.5-mmproj.gguf | multimodal projector (vision) |
| experts-htp-tiled/layer_000.bin .. layer_039.bin | 40 layers of Q4_K/Q6_K experts, Hexagon tiled layout |
| experts-htp-tiled/manifest.json | expert metadata (`backend_layout = hexagon-tiled-v1`) |
| SHA256SUMS | file integrity |
| run-sage-text-fast.sh | example text run script |

Prepacked experts require the matching Hexagon backend. Set
`LLAMA_MOE_HTP_PACKED_DIR` and `--moe-experts` both to `experts-htp-tiled`.
Pass `--mmproj models/Sage-A1.5-mmproj.gguf` for image input. Processing is
extraction/repacking only; no retraining. Q4_K_M expert quantization; accuracy
equivalence is not established. Upstream Apache-2.0 licenses apply.