Document RKLLM artifacts and checksums

README.md

@@ -15,32 +15,31 @@ tags:
 
 # Qwen3.5-4B-RKLLM
 
-RKLLM-converted Qwen3.5-4B artifacts for deployment on Rockchip RK3576 and RK3588 NPUs. This repository includes language-model binaries and the matching vision encoder artifacts for multimodal inference.
+RKLLM/RKNN-converted Qwen3.5-4B multimodal artifacts for Rockchip RK3576 and RK3588 NPUs.
 
-`.rkllm` and `.rknn` files require the Rockchip RKLLM/RKNN runtime. They cannot be loaded directly with Transformers, llama.cpp, or Ollama.
+This is a VLM conversion: each supported platform requires both the `.rkllm` language model and the matching `.rknn` vision encoder. The pair must come from the same platform directory. These are hardware-specific artifacts, not Transformers checkpoints.
 
-## Base Model
+## Base model
 
-- Model: Qwen3.5-4B
-- Author: Qwen Team
-- Original model: [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B)
-- Original license: Apache-2.0
+- Upstream model: [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B)
+- License: Apache-2.0
+- Model type: VLM (vision-language model)
 
-Refer to the upstream model card for the original model's capabilities, limitations, and acceptable-use guidance.
+## Conversion and variants
 
-## Available Artifacts
+Converted with RKLLM Toolkit 1.3.0 and an RKNN vision conversion. Use a matching pair for the exact target SoC.
 
-| Target | Language model | Quantization | Vision model |
-| --- | --- | --- | --- |
-| RK3576 | [`Qwen3.5-4B_RK3576_w4a16_g128.rkllm`](RK3576/Qwen3.5-4B_RK3576_w4a16_g128.rkllm) | W4A16 (g128) | [`Qwen3.5-4B_vision_RK3576.rknn`](RK3576/Qwen3.5-4B_vision_RK3576.rknn) |
-| RK3576 | [`Qwen3.5-4B_RK3576_w8a8.rkllm`](RK3576/Qwen3.5-4B_RK3576_w8a8.rkllm) | W8A8 | [`Qwen3.5-4B_vision_RK3576.rknn`](RK3576/Qwen3.5-4B_vision_RK3576.rknn) |
-| RK3588 | [`Qwen3.5-4B_RK3588_w8a8.rkllm`](RK3588/Qwen3.5-4B_RK3588_w8a8.rkllm) | W8A8 | [`Qwen3.5-4B_vision_RK3588.rknn`](RK3588/Qwen3.5-4B_vision_RK3588.rknn) |
+| Target | Quantization | RKLLM language model | RKLLM SHA256 | RKNN vision encoder | RKNN SHA256 |
+| --- | --- | --- | --- | --- | --- |
+| RK3576 | W4A16 (g128) | [Qwen3.5-4B_RK3576_w4a16_g128.rkllm](RK3576/Qwen3.5-4B_RK3576_w4a16_g128.rkllm) | `aa4d34b42752a0e491ed891ce4b7f32a745631601a65b175f743f971bdb33482` | [Qwen3.5-4B_vision_RK3576.rknn](RK3576/Qwen3.5-4B_vision_RK3576.rknn) | `6692f52adaa7ee0ba892c7cbba01979750cb82cff8061a571f56e9d84c6d98a2` |
+| RK3576 | W8A8 | [Qwen3.5-4B_RK3576_w8a8.rkllm](RK3576/Qwen3.5-4B_RK3576_w8a8.rkllm) | `736f4b1065b481bae9cbd86e6c9ed30222abba9c5a3cb2d8d68dd92037d82dfe` | [Qwen3.5-4B_vision_RK3576.rknn](RK3576/Qwen3.5-4B_vision_RK3576.rknn) | `6692f52adaa7ee0ba892c7cbba01979750cb82cff8061a571f56e9d84c6d98a2` |
+| RK3588 | W8A8 | [Qwen3.5-4B_RK3588_w8a8.rkllm](RK3588/Qwen3.5-4B_RK3588_w8a8.rkllm) | `715566bbee72b25d8c4912f6cf6256a8ac264a754511573a99a541487bbc06b4` | [Qwen3.5-4B_vision_RK3588.rknn](RK3588/Qwen3.5-4B_vision_RK3588.rknn) | `c286ef69266c11a2a2cedd881ad6a40415c17c8924962f07454359bb6d60e2ba` |
 
-The repository also contains `Qwen3.5-4B_vision.onnx`, the vision encoder in ONNX format, and `Qwen3.5-4B_data_quant.json`, the calibration data used during conversion.
+The root `Qwen3.5-4B_vision.onnx` is the vision conversion input; use the platform-specific `.rknn` encoder for deployment.
 
 ## Usage
 
-Download one language-model artifact and the matching vision artifact for your SoC. Do not mix RK3576 and RK3588 files.
+Download both files for the target platform:
 
 ```bash
 hf download HanzoHuang/Qwen3.5-4B-RKLLM \
@@ -49,14 +48,12 @@ hf download HanzoHuang/Qwen3.5-4B-RKLLM \
   --local-dir Qwen3.5-4B-RKLLM
 ```
 
-Load the `.rkllm` model with a compatible RKLLM runtime and the `.rknn` vision encoder with the matching RKNN runtime. Application code must implement Qwen3.5's multimodal preprocessing and prompt format.
+Use them with the RKLLM VLM runtime. For a Docker API, see [Hanzo-Huang/rkllm-docker](https://github.com/Hanzo-Huang/rkllm-docker) and set `MODEL_KIND=vlm` with both model files.
 
 ## Limitations
 
-- These are hardware-specific converted artifacts, not Transformers checkpoints.
-- Runtime, driver, and toolkit compatibility depends on the Rockchip software stack installed on the device.
-- Conversion may change output quality relative to the upstream floating-point model; validate on your own workload.
+The vision encoder and language model are SoC-specific and must be kept as a matching pair. Validate image preprocessing, memory use, and runtime compatibility on your device.
 
 ## Acknowledgements
 
-Thanks to the Qwen Team for releasing Qwen3.5-4B and to Rockchip and RKLLM contributors for the deployment toolchain.
+Thanks to the Qwen Team, Rockchip, and the RKLLM/RKNN community.