LightNav-0-GGUF

LightNav-0-GGUF

LightNav-0 is a compact generalist embodied navigation model built on Qwen3-VL-4B-Instruct that elicits the pretrained vision-language model's spatial intelligence and aligns it with navigation tasks without task-specific prediction heads. Its core design unifies diverse tasks — instruction following, open-vocabulary object navigation, and visual tracking — under a single token interface: dual-channel pointing tokens (<apos_*>/<opos_*>) express task-, scene-, and embodiment-agnostic spatial intent, which a residual vector-quantized (RVQ) action tokenizer then maps into precise, embodiment-specific trajectories via three levels of action tokens (<act_l0_*><act_l1_*><act_l2_*>) decoded into a 10-step chunk of robot-local waypoints (forward distance, lateral offset, yaw). The model takes a SlowFast-compressed history of first-person RGB frames (256×448 at 4fps) plus a natural-language instruction as input, loads directly with stock Transformers since navigation tokens are ordinary embedding-table rows, and is served through the companion lightnav package via vLLM for offline prediction or WebSocket-based real-time tracking/navigation serving, with Habitat VLN-CE/ObjectNav and EVT-Bench evaluation recipes plus real-robot deployment guides documented in its GitHub repository. It is released under the Apache License 2.0.

RVQ action-tokenizer bundle shared by navigation and tracking (one 3 × 256 residual codebook, horizon 10) — https://huggingface.co/LightOriginsHQ/LightNav-0/tree/main/action_tokenizer

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp