Community
Open-source RK3588 NPU runtime runs Qwen3.5-4B at about 10 tokens per second
The open_rknpu2_lm_infer engine runs without Rockchip's proprietary userspace libraries and reports roughly 2x faster decode than RKLLM 1.3.1 on the RK3588 NPU. It supports image input and MTP speculative decoding.