Qwen3.5-0.8B

Qwen3.5-0.8B for SpacemiT K1/K3

This is the SpacemiT deployment package for 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 and cite the original work:

@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 and an SMT-enabled SpacemiT llama.cpp. Prebuilt packages can be unpacked directly:

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

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.