dtaddis/LFM2.5-VL-450M-Unity-Sentis

LFM2.5-VL-450M — Unity Sentis UINT8 graphs

Unity AI Inference (Sentis) versions of Liquid AI's LFM2.5-VL-450M model, prepared for fully offline inference in Unity 6 on desktop, Android, and iOS.

This is an independent derivative repository, not an official Liquid AI release. The three model files were converted from Liquid AI's FP16 ONNX release with Unity AI Inference 2.5.0's built-in ModelQuantizer using QuantizationType.Uint8, then saved in Unity's native .sentis format. No model architecture or tokenizer behavior was intentionally changed.

Files

File	Purpose	Approximate size
embed_tokens_uint8.sentis	Token embedding graph	67 MB
vision_encoder_uint8.sentis	Vision encoder graph	95 MB
decoder_model_merged_uint8.sentis	Autoregressive decoder and cache graph	363 MB
tokenizer.json	Original Liquid AI tokenizer	4.7 MB

The full download is approximately 530 MB. These graphs avoid the unsupported MatMulNBits operators in the upstream Q4 ONNX decoder and import directly as Unity.InferenceEngine.ModelAsset files.

Unity usage

These files are the tested download set for Interactive Offline Vision LLM and its source repository. Put all files in Assets/Models/LFM2.5-VL-450M, assign the three .sentis assets and tokenizer.json to the InteractiveOfflineVisionLLM component, and use the GPUCompute backend on supported devices.

The Unity asset's Editor downloader performs the initial download. Built players embed the referenced files and do not need a network connection for inference.

Attribution and licence

• Original model: LiquidAI/LFM2.5-VL-450M
• Original ONNX export: LiquidAI/LFM2.5-VL-450M-ONNX
• Copyright and model authorship: Liquid AI, Inc.

The model, tokenizer, and these derivative weight files are governed by the LFM Open License v1.0. Read LICENSE before downloading or using them. See MODEL_NOTICE.md for the exact modifications and source/output hashes.