Qwen3-4B-Instruct-2507 model resources

Minimal inference code and deployment steps: https://github.com/APLUX-Agora/qwen3-4b-instruct-demo

Targets: QCS8550, QCS8625, IQ8 and IQ9. Compiled context: 4096 tokens, not the source model's 256K context. Quantization: W4A16. Model binaries were compiled with QNN 2.36; the example uses the installed AidGen backend.

Download

export AGORA_WORKSPACE="$HOME/my-qwen-demo"
export AGORA_SOC='QCS8550'
hf download APLUX-Agora/Qwen3-4B-instruct-2507 \
  --include "targets/$AGORA_SOC/*" --local-dir "$AGORA_WORKSPACE/resources"

Set the two global variables directly in your device terminal; choose any workspace path. The GitHub README provides the same five deployment steps. Each targets/<SoC>/ contains three model shards, tokenizer.json, config.json and htp_backend_ext_config.json. Keep these files together. About 3.2 GiB per target. The minimal example runs without a prefix KV cache.

Model and tokenizer bytes are unchanged; paths and relative configuration references are standardized for Agora. Source model: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507.

Source model license: Apache-2.0 (LICENSE_SOURCE). Converted resources: APLUX Model Farm License (LICENSE_MODEL_FARM.pdf). See NOTICE.md and SHA256SUMS for provenance and binary checksums.