license: other
license_name: lfm-open-license-v1.0
license_link: https://huggingface.co/LiquidAI/LFM2.5-350M/blob/9921cd4d82c0e3986887aa4b376f9446729ef8a3/LICENSE
base_model:
- PinoCookie/LFM2.5-350M-abliterated
- PinoCookie/LFM2.5-1.2B-JP-Abliterated
library_name: onnxruntime
tags: - onnx
- lfm2
- lfm2.5
- quantized
- abliterated
HINA browser Q4F16 conversions of community LFM2.5 derivatives
These are experimental browser conversions, not official Liquid AI releases.
| Folder | Community source | Pinned source revision |
|---|---|---|
| 350m | PinoCookie/LFM2.5-350M-abliterated | e1f9ece446c1a50fe44afa1709d0cfba518151a5 |
| 1.2b-jp | PinoCookie/LFM2.5-1.2B-JP-Abliterated | ebba6787ed43d8bbbee0d531e757a8d117de82cb |
All 148 learned tensors of each model come from its community checkpoint. Matrices are quantized to block-32 4-bit weights with FP16 scales; other learned tensors are FP16. Graph topology comes from the HINA app's pinned official LiquidAI ONNX graphs, with a last-position-only LM head. The common rotary tables use head dimension 64, theta 1,000,000, and 128,000 positions. Source hashes, graph hashes, tensor coverage and artifact SHA-256 values are in each folder's provenance.json. The reproducible conversion script is conversion.py.
Use ONNX Runtime Web with WebGPU and shader-f16 support. The browser application limits context to 8,192 tokens and waits for GPU completion between prefill chunks. Original Safetensors checkpoints are not included.
Attribution and terms
Credit: Liquid AI for the base models and ONNX graphs; PinoCookie for the community checkpoints. Source model cards are preserved as SOURCE-MODEL-CARD.md in each folder. Base model LICENSE texts are preserved in each folder. The upstream Japanese derivative's card advertises Apache-2.0, but this conversion preserves the actual LiquidAI base model license rather than treating that metadata as a replacement for upstream terms.
Quality
These community checkpoints modify refusal behavior. They are not guaranteed to answer every prompt and the conversion is not a quality certification. Japanese accuracy, reasoning quality, repetition, runtime stability and memory use require application-specific testing. These are experimental weights, not a promise of production suitability.