Qwen3.5-4B for SpacemiT K1/K3

This is a SpacemiT conversion of Qwen/Qwen3.5-4B, the Apache-2.0 native multimodal model from the Qwen Team. It is intended for visual understanding, text, coding and agent tasks. Read the official announcement 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}}

Contents: qwen3_5_4b-text-q41.gguf, ONNX vision encoders at 224/384/768 pixels (384 is the default), and independent K1/K3 configs. The text model is Q4_1 GGUF; vision runs through the SpaceMIT Execution Provider.

Prerequisites are the SpacemiT ONNX Runtime releases and SMT-enabled spacemit-com/llama.cpp. Prebuilt ORT 2.0.6 and llama.cpp v0.1.7 archives can be unpacked directly. To compile, clone recursively, set RISCV_ROOT_PATH and SPACEMIT_ORT_DIR, and run bash build_spacemit.sh glibc.

K1 uses AI cores 0–3, configs/K1, and -t 4; K3 uses AI cores 8–15, configs/K3, and -t 8. Never swap these configs: ep_config pins the vision EP to the platform's AI cores.

export MODEL_DIR=/path/to/Qwen3.5-4B-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 "$MODEL_DIR/qwen3_5_4b-text-q41.gguf" --media-backend smt --smt-config-dir "$MODEL_DIR/configs/K1" -t 4 --host 0.0.0.0 --port 8080 --warmup

For K3 select configs/K3 and -t 8. Send an image data URL plus Describe the image content. to /v1/chat/completions (64 output tokens, temperature 0, thinking disabled). The bundled smoke image produced HTTP 200 on K1 (description beginning This image captures a moment from what appears to be a formal event...) and K3 (beginning This image captures a moment from what appears to be a conference...).

The Qwen3.5 weights are Apache-2.0. Runtime dependencies remain under their upstream licenses.