Access LVL-1
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
By requesting access, you agree to share your Hugging Face username and email address with Vitrus and to use LVL-1 only under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0).
Log in or Sign Up to review the conditions and access this model content.
Leãozinho VL-1 (LVL-1) ONNX
ONNX component graphs for the native vitrus/LVL-1 checkpoint at step 6000.
The FP32 reference export consists of the DINOv2 detail encoder, fixed-4:3 SigLIP2 NaFlex semantic encoder, dual visual projector, cached decoder prefill, and cached one-token decoder. The decoder graphs preserve LVL-1's grouped-query attention by expanding its two KV heads to eight query heads in the exported graph, which is mathematically equivalent.
Status
This release contains verified FP32 and FP16 component graphs and a validated dynamic per-channel INT8-weight decoder profile. The INT8 profile leaves vision and the projector at FP16/FP32, so it is not an all-INT8 VLM. INT4 and the browser WebGPU runtime are added only after their own accuracy and device tests pass. Do not interpret the presence of a graph as an unqualified browser-performance claim.
The initial SigLIP2 graph is pinned to the 4:3 / 14×18 NaFlex profile used by Vitrus' curated browser examples. Native LVL-1 keeps dynamic-aspect preprocessing. A dynamic-aspect browser graph is not yet claimed.
License and limitations
The files are licensed under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0). You may share and adapt them for noncommercial purposes with attribution. Commercial use requires separate permission from Vitrus. LVL-1 can hallucinate or repeat scene details, and its current 2D-box mode can emit full-frame boxes. It must not be used as a safety or robot-control authority.
Browser demo
The reference browser demo is in webgpu-demo/ in the release source tree. It streams these public graphs to ONNX Runtime Web, attempts WebGPU first, and explicitly reports any WASM fallback. It is the fixed 4:3 SigLIP profile above; no device-performance result is implied until a browser run reports it.