Qwen3.5-2B for SpacemiT K1/K3

This package deploys Qwen/Qwen3.5-2B, the Qwen Team's Apache-2.0 native multimodal vision-language model. It supports visual understanding, text, coding and agent workflows. See the Qwen3.5 blog and cite:

@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 Q4_1 GGUF text decoder is paired with ONNX vision encoders at 224, 384 (default), and 768 pixels. Files are qwen3_5_2b-text-q41.gguf, the three qwen3_5_2b-vision-*-op23.f16.onnx files, and configs/K1 / configs/K3.

Install SpacemiT ONNX Runtime and SMT-enabled llama.cpp, or unpack spacemit-ort.riscv64.2.0.6.tar.gz and spacemit-llama.cpp.riscv64.0.1.7.tar.gz. Source builds use git clone --recursive, RISCV_ROOT_PATH, SPACEMIT_ORT_DIR, and bash build_spacemit.sh glibc.

K1/X60 has AI cores 0–3: use configs/K1 and -t 4. K3/A100 has AI cores 8–15: use configs/K3 and -t 8. The configs' ep_config contains the corresponding affinity.

export MODEL_DIR=/path/to/Qwen3.5-2B-SpacemiT LLAMA_DIR=/path/to/llama.cpp-installed ORT_DIR=/path/to/spacemit-ort.riscv64.2.0.6
export LD_LIBRARY_PATH="$LLAMA_DIR/lib:$ORT_DIR/lib:${LD_LIBRARY_PATH:-}"
"$LLAMA_DIR/bin/llama-server" -m "$MODEL_DIR/qwen3_5_2b-text-q41.gguf" --media-backend smt --smt-config-dir "$MODEL_DIR/configs/K1" -t 4 --host 0.0.0.0 --port 8080 --warmup

For K3 change K1 to K3 and -t 4 to -t 8. POST an image data URL and Describe the image content. to /v1/chat/completions with max_tokens: 64, temperature: 0, and thinking disabled. Board smoke tests with humanspeech.jpg returned HTTP 200 on both K1 (0-3, output beginning Here's a detailed description of the image:) and K3 (8-15, output beginning The image displays a collection of books...).

The original model is Apache-2.0; llama.cpp and ONNX Runtime retain their own licenses.