HanzoHuang/Qwen2.5-VL-3B-Instruct-RKLLM

Qwen2.5-VL-3B-Instruct · RKLLM

Text and image chat on Rockchip RK3576 and RK3588, converted from Qwen/Qwen2.5-VL-3B-Instruct.

• Max context length: 8192

• Conversion toolkit: RKLLM Toolkit 1.3.1 and RKNN Toolkit2 2.5.0.
• Inputs / output: text and one image → text. Video and audio input are unavailable in this runtime.
• License: Apache 2.0; see the original model and license.

Choose your files

Use files for your board's exact SoC. RK3576 offers compact W4A16-g128 and W8A8; RK3588 uses W8A8. Download size is not runtime RAM.

Board	Quantization	Size	Language model
RK3576	W4A16-g128	2.60 GiB	Qwen2.5-VL-3B-Instruct_RK3576_w4a16_g128.rkllm
RK3576	w8a8	3.70 GiB	Qwen2.5-VL-3B-Instruct_RK3576_w8a8.rkllm
RK3588	w8a8	3.51 GiB	Qwen2.5-VL-3B-Instruct_RK3588_w8a8.rkllm

Every language model requires its board's matching vision encoder below. Both RK3576 language variants share the RK3576 encoder.

Board	Size	Vision encoder
RK3576	1.52 GiB	Qwen2.5-VL-3B-Instruct_vision_RK3576.rknn
RK3588	1.38 GiB	Qwen2.5-VL-3B-Instruct_vision_RK3588.rknn

Run on your board

The vision encoders require RKNN Runtime 2.5.0 and the language models require RKLLM Runtime 1.3.1. Set RKLLM_RUNTIME_IMAGE to an ARM64 runtime image with those versions. The validation used a locally built image; a model-specific Docker image is not included in this release.

Download the RK3576 W4A16-g128 language model and matching encoder:

hf download HanzoHuang/Qwen2.5-VL-3B-Instruct-RKLLM \
  v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_RK3576_w4a16_g128.rkllm \
  v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_vision_RK3576.rknn \
  --local-dir ./models

sudo docker run --rm -it --privileged \
  -p 8001:8001 -v /dev:/dev \
  -v "$PWD/models/v1.3.1/RK3576:/app/models:ro" \
  -e MODEL_KIND=vlm -e TARGET_PLATFORM=rk3576 -e RKLLM_EMBED_FLASH=1 \
  -e MODEL_FILE=Qwen2.5-VL-3B-Instruct_RK3576_w4a16_g128.rkllm \
  -e VISION_MODEL_FILE=Qwen2.5-VL-3B-Instruct_vision_RK3576.rknn \
  ${RKLLM_RUNTIME_IMAGE:?Set RKLLM_RUNTIME_IMAGE to an ARM64 RKLLM 1.3.1 / RKNN 2.5.0 image} \
  serve --max_context_len 4096

For RK3588, download both RK3588 files, mount that directory, set TARGET_PLATFORM=rk3588, and select its W8A8 language filename. Use the runtime documentation for health checks and OpenAI-compatible image requests.

Release notes

• Source acquired from ModelScope Qwen/Qwen2.5-VL-3B-Instruct, pinned revision 1b5a7674d40a0ac9acef541a46fed12d710bd36b. The HF link above identifies the canonical upstream model; no equivalence to an HF Git commit is assumed.
• Export context: 8192 tokens. Functional API tests used 4096; the speed/memory benchmark used 2048. Longer runtime context has not been validated on these boards.
• Vision export follows the SDK placement rule, including a CPU-targeted patch-embedding convolution.
• Fixed vision input: RGB, 448 × 448, batch 1, 256 image tokens, CLIP mean [0.48145466, 0.4578275, 0.40821073] and std [0.26862954, 0.26130258, 0.27577711]. The runtime pads to a square using RGB 127, then resizes bilinearly. The encoder returns one pooler output with 2048 channels per image token.
• Language calibration: the SDK's 20 image-and-text examples, captured using the source processor and language-model input hook. W4A16-g128 uses GRQ; W8A8 uses normal quantization. Float32 CUDA loading; two NPU cores on RK3576 and three on RK3588.
• Export compatibility uses the pooler output under Transformers 5, imports only the requested calibration model family, and supplies a temporary nested text_config view for the SDK language loader. It restores the source mrope_section rotary configuration after SDK model loading. Original checkpoint weights and configuration files are unchanged.
• All three variants passed seven API smoke checks on their target board: arithmetic, independent requests, SSE, exact conversation recall, image description, image color over SSE, and text after an image. Tests used the baked RKLLM 1.3.1 application with RKNN Runtime 2.5.0 and embed_flash=1.
• RK3576 vision encoder cosine against original float32 embeddings: 0.99772957.
• RK3588 vision encoder cosine against original float32 embeddings: 0.99708773.

These are functional and numerical smoke checks, not a broad quality benchmark. Quantization and fixed image resolution may reduce accuracy.

Measured on the target boards:

Board	Quantization	Decode tok/s	Startup RAM (GB)	Stable RAM (GB)	Date
RK3576	w4a16-g128	6.17	4.19	4.14	2026-10-23
RK3576	w8a8	5.13	5.32	5.32	2026-10-23
RK3588	w8a8	8.47	4.97	5.00	2026-10-23

Speed is the median native decode rate over three 128-token runs after a 32-token warmup, greedy decoding, EOS ignored, context 2048. RAM is additional board usage above a five-second idle baseline: sampled loading peak and ten-second ready-state median, in decimal GB. The matched encoder is loaded; embed_flash=1. First-token timing stays in the private report. Measurements depend on the board, clocks, cooling, and background load.

SHA-256 checksums

de7152bfb6c2f532a048ac0e50a353a7a74dc8f10009af8f99e80fb273eda830  v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_RK3576_w4a16_g128.rkllm
41f1488381f16666ce89545028f9ec8457052be591c036cd68a309af7d671612  v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_RK3576_w8a8.rkllm
e5bb20a18703232c14870c0abb157950e5b98f3c78b9c275a1c602c17efd85c3  v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_vision_RK3576.rknn
3621ae4e255f08fba6b7b4a46ccfda3775bd4b8da6cdf4336aeff37947b8eba3  v1.3.1/RK3588/Qwen2.5-VL-3B-Instruct_RK3588_w8a8.rkllm
bc6620d506b583b2c4bbd51fea9aecdf07d109a5cbbb5e7e35906de2041e009a  v1.3.1/RK3588/Qwen2.5-VL-3B-Instruct_vision_RK3588.rknn

Credits

Qwen provides the original model. Conversion uses Rockchip RKLLM, RKNN Toolkit2, and the conversion wrapper.