File size: 3,015 Bytes
d5092bf 4e16011 d5092bf 4e16011 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 | ---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: llama.cpp
tags: [qwen3.5, spacemit, k1, k3, gguf, onnxruntime]
---
# Qwen3.5-0.8B for SpacemiT K1/K3
This is the SpacemiT deployment package for [Qwen/Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B), an Apache-2.0 native multimodal vision-language model for image understanding, text, coding and agent tasks. Please see the Qwen Team's [Qwen3.5 announcement](https://qwen.ai/blog?id=qwen3.5) and cite the original work:
```bibtex
@misc{qwen3.5, title={{Qwen3.5}: Towards Native Multimodal Agents}, author={{Qwen Team}}, month={February}, year={2026}, url={https://qwen.ai/blog?id=qwen3.5}}
```
The text decoder is Q4_1 GGUF and the vision encoder is ONNX. The default configuration uses the 384-pixel encoder; 224- and 768-pixel alternatives are included.
## Files and platforms
`qwen3_5vl_0.8b-text-q41.gguf`, three `qwen3_5vl_0.8b-vision-*-op23.f16.onnx` files, and `configs/K1/config.json` / `configs/K3/config.json` are included. K1/X60 uses AI cores 0–3 and `-t 4`; K3/A100 uses AI cores 8–15 and `-t 8`. Use the matching configuration: its `ep_config` sets the SpaceMIT EP affinity.
## Prerequisites
Install the [SpacemiT ONNX Runtime release](https://github.com/spacemit-com/onnxruntime/releases) and an SMT-enabled [SpacemiT llama.cpp](https://github.com/spacemit-com/llama.cpp). Prebuilt packages can be unpacked directly:
```bash
wget https://github.com/spacemit-com/onnxruntime/releases/download/2.0.6/spacemit-ort.riscv64.2.0.6.tar.gz
wget https://github.com/spacemit-com/llama.cpp/releases/download/v0.1.7/spacemit-llama.cpp.riscv64.0.1.7.tar.gz
tar -xf spacemit-ort.riscv64.2.0.6.tar.gz; tar -xf spacemit-llama.cpp.riscv64.0.1.7.tar.gz
```
To build llama.cpp, clone recursively, set `RISCV_ROOT_PATH` and `SPACEMIT_ORT_DIR`, then run `bash build_spacemit.sh glibc`.
## Run
```bash
export MODEL_DIR=/path/to/Qwen3.5-0.8B-SpacemiT
export LLAMA_DIR=/path/to/llama.cpp-installed
export ORT_DIR=/path/to/spacemit-ort.riscv64.2.0.6
export LD_LIBRARY_PATH="$LLAMA_DIR/lib:$ORT_DIR/lib:${LD_LIBRARY_PATH:-}"
```
K1: `"$LLAMA_DIR/bin/llama-server" -m "$MODEL_DIR/qwen3_5vl_0.8b-text-q41.gguf" --media-backend smt --smt-config-dir "$MODEL_DIR/configs/K1" -t 4 --host 0.0.0.0 --port 8080 --warmup`
K3: use the same command with `configs/K3` and `-t 8`.
Send an OpenAI-compatible request to `/v1/chat/completions` with an `image_url` data URL and text such as `Describe the image content.`, `max_tokens: 64`, `temperature: 0`, and `chat_template_kwargs: {"enable_thinking": false}`.
## Board verification
Using `humanspeech.jpg`, ORT 2.0.6 and SMT llama.cpp, both boards returned HTTP 200. K1 began `The image captures a lively scene of a speech or presentation...`; K3 began `This image captures a lively scene, likely from a formal event or presentation...`. These are functional smoke tests, not benchmarks.
The original Qwen3.5 model is Apache-2.0; dependency licenses remain with their respective projects.
|