parkingsoman/LFM2.5-VL-3B-q4f16-webgpu

LFM2.5-VL-3B q4f16 decoder for WebGPU (modified files)

MODIFIED FILES. The files in this repository are not the files that Liquid AI published. They were changed from LiquidAI/LFM2.5-VL-3B-ONNX at revision 020f9311343b8f3c6d9789b4dd300e749ec52cd1. Liquid AI did not make or review these changes.

LFM2.5-VL-3B is a model by Liquid AI. The original model card is at https://huggingface.co/LiquidAI/LFM2.5-VL-3B. This repository is distributed under the LFM Open License v1.0 (see LICENSE, a copy of the licence file in LiquidAI/LFM2.5-VL-3B-ONNX) and carries its NOTICE.

What was changed

This repository holds only the decoder. Every other file that the model needs (vision encoder, token embeddings, tokenizer, configuration, chat template) is unchanged and is fetched from LiquidAI/LFM2.5-VL-3B-ONNX at revision 020f9311343b8f3c6d9789b4dd300e749ec52cd1.

Change	Detail
Source file	onnx/decoder_model_merged_q4.onnx and its external data, from LiquidAI/LFM2.5-VL-3B-ONNX at 020f9311343b8f3c6d9789b4dd300e749ec52cd1
Re-quantised	activations and non-quantised weights converted from fp32 to fp16 (q4f16); the int4 weights are unchanged
lm_head repaired	the tied output layer (model.embed_tokens.weight) kept in fp32, and the Cast that made it fp16 removed
Re-sharded	external data split into 4 files; none is 2 GiB or larger

The converter is browser-prototype/scripts/q4f16/convert-q4f16-lmhead-repair.py in the Axiya repository.

Files

File	Bytes	SHA-256
onnx/decoder_model_merged_q4f16.onnx	338513	1d7af7fbf0fe5d31acc988a7a182f4c5bec46071bf5a42cce65390f207ecff09
onnx/decoder_model_merged_q4f16.onnx_data	16384000	69e5517c008545487cb644cd85484e650e68cb1308ea84947b3ad972317db8d4
onnx/decoder_model_merged_q4f16.onnx_data_1	1048576000	7454498e9aade5f3767cb3d62f91a72fb76bca552979fba3972b9c65cd1a1c8e
onnx/decoder_model_merged_q4f16.onnx_data_2	992014336	9296100fec8349c2eda4ef225441ba34a7881b7f80564066808d3b71ab767b7c
onnx/decoder_model_merged_q4f16.onnx_data_3	378077184	f0fff61384f70ea44c773c943841b3d726b17816a9677e9ae83e390408c87153

Use

These files are pinned by content hash in the Axiya browser app (browser-prototype/src/model-lock.q4f16.json, build flag VITE_MODEL=q4f16). The app loads them with Transformers.js and ONNX Runtime Web on WebGPU, with this dtype map:

{ embed_tokens: "fp16", decoder_model_merged: "q4f16", vision_encoder: "fp16" }

The app downloads each file from this repository at a pinned commit and checks its SHA-256 before it stores it.