Qwen-Drive-1.0-4B GGUF
This repository packages the released Qwen-Drive-1.0-4B checkpoints as GGUF components for the native rust-model-inference runtime. It is a conversion, not a new model or a quantization: the main VLM and both planners retain BF16 payloads where supported; required norms, biases, and vision-patch tensors are F32. The perception head is F32.
Qwen-Drive-1.0 combines a Qwen3.5-4B vision-language model with external 3D perception and motion-planning heads. Read the official implementation and technical report for the model, data, evaluation, and intended-use details.
Files
All five files are required for full functionality. Choose exactly one planner.
| File | Size | Role |
|---|---|---|
Qwen-Drive-1.0-4B-BF16.gguf | 7.85 GiB | Qwen3.5 vision-language model |
Qwen-Drive-1.0-4B-mmproj-BF16.gguf | 629 MiB | Vision projector; used with the VLM |
Qwen-Drive-1.0-planner-sft-BF16.gguf | 1.94 GiB | Supervised-finetuned planning expert |
Qwen-Drive-1.0-planner-rl-BF16.gguf | 1.94 GiB | Reward-optimized planning expert |
Qwen-Drive-1.0-perception-F32.gguf | 477 MiB | BEV 3D perception head |
Integrity
The bundled exporter verified each file's GGUF directory and SHA-256 digest.
| File | SHA-256 |
|---|---|
Qwen-Drive-1.0-4B-BF16.gguf | 1d5aabe7f02ef97fdf173bad8bcaaf783af4d85bd07ea56cc3dd42d1f28deac9 |
Qwen-Drive-1.0-4B-mmproj-BF16.gguf | d0ba72870cca4073c0e3aaef251e36c6ff1061c6b76fa725cb55d518a11a3d9f |
Qwen-Drive-1.0-planner-sft-BF16.gguf | 7ab6abef07523192aad877bc0186dcc6c82c854f951f00ba00c7ce06bcf89cd9 |
Qwen-Drive-1.0-planner-rl-BF16.gguf | 05f205728c822382650b6ee43f8c343da09f2dce381e399c86f891b4411b9b5a |
Qwen-Drive-1.0-perception-F32.gguf | c1ac01ca2f3c70a67d2517969d43016373c4b30dca47ce82b29609739b9879c5 |
Usage
Build the runtime, download the required components, and run planning with either the SFT or RL planner:
cargo build --release --bin rust-model-inference
target/release/rust-model-inference \
--model Qwen-Drive-1.0-4B-BF16.gguf \
--mmproj Qwen-Drive-1.0-4B-mmproj-BF16.gguf \
--planner Qwen-Drive-1.0-planner-sft-BF16.gguf \
--scenes /path/to/planning_scenes.jsonl \
--image-root /path/to/images \
--mode direct_planning \
--num-samples 1 --num-steps 10 --seed 42 \
--output predictions.jsonl --threads 12
Perception additionally requires the VLM, mmproj, perception component, six camera JPEGs, and the official-format frame manifest:
target/release/rust-model-inference \
--model Qwen-Drive-1.0-4B-BF16.gguf \
--mmproj Qwen-Drive-1.0-4B-mmproj-BF16.gguf \
--perception Qwen-Drive-1.0-perception-F32.gguf \
--frames /path/to/frame \
--output perception.json --threads 12
For frame preparation, input schemas, planning modes, and limitations, follow the official repository. The GGUF exporter and runtime examples are in tools/qwen_drive.
License and responsible use
This conversion preserves the upstream Apache-2.0 license. Follow the upstream model's terms and guidance. This is research software for offline evaluation and is not a vehicle-control system; do not use it to operate a vehicle or make safety-critical decisions without independent validation and appropriate safeguards.