GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
LightOriginsHQ
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
LightOriginsHQ
Model Tree
Qwen/Qwen3-VL-4B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| Path | Content |
|---|---|
config.json, model-*.safetensors, model.safetensors.index.json | Qwen3-VL weights (bf16) with the navigation token embeddings |
tokenizer.json, tokenizer_config.json, chat_template.jinja, processor_config.json | tokenizer / processor |
eval_config.json | processing parameters that must match training (video size, pooling, SlowFast history tiers, task settings); read automatically by the inference code |
action_tokenizer/ | RVQ action-tokenizer bundle shared by navigation and tracking (one 3 × 256 residual codebook, horizon 10) |
The bundle holds manifest.json, the per-level codebooks (codebook_l{0,1,2}.npy, 3 × 256 codes),
jacobian_weights.npy and alpha_per_source.json. The model emits one <act_l{level}_{code}>
token per level; the bundle decodes them into a (10, 3) chunk of robot-local waypoints
[forward_m, lateral_m (+left), yaw_rad (+ccw)]. eval_config.json references the bundle by a
path relative to this directory, so the inference code finds them without extra flags.
bash
pip install "lightnav[vllm,video] @ git+https://github.com/lightrobo/LightNav-0"hf download LightOriginsHQ/LightNav-0 --local-dir ./LightNav-0-ckpt# offline prediction on a cliplightnav-predict --model_path ./LightNav-0-ckpt --backend vllm_local \--video clip.mp4 --fps 4 --instruction "follow the person in the red shirt"# serve (tracking prompt) and drive with the reference clientlightnav-serve --task tracking --model_path ./LightNav-0-ckpt --backend vllm_local --port 8050lightnav-ws-client --server ws://localhost:8050 --video clip.mp4 --fps 4 \--instruction "follow the person in the red shirt"# navigation prompt (instruction following / object navigation)lightnav-serve --task vln --model_path ./LightNav-0-ckpt --backend vllm_local --port 8051
Habitat VLN-CE / ObjectNav and EVT-Bench evaluation recipes, the WebSocket protocol and the
real-robot deployment guide are in the code repository's docs/.
transformers.<apos_*> / <opos_*>) followed by three RVQ action
tokens (<act_l0_*><act_l1_*><act_l2_*>) decoded to a 10-step waypoint chunk.Questions, deployment notes and release news — join us on Discord, or scan to join the WeChat group:
Apache License 2.0.