LFM2.5-2.6B-Uncensored-ONNX

LFM2.5-2.6B-Uncensored ONNX

ONNX Runtime / Transformers.js exports of SC117/LFM2.5-2.6B-Uncensored. These files are intended for browser and cross-platform inference, especially WebGPU.

This repository contains the two web-oriented variants:

Q4F16 is not a GGUF quantization. It uses INT4 weights with FP16 runtime tensors and KV cache. It is therefore not interchangeable with Q4_K_M.

Transformers.js

npm install @huggingface/transformers

import { pipeline, TextStreamer } from "@huggingface/transformers";

const generator = await pipeline(
  "text-generation",
  "acidsound/LFM2.5-2.6B-Uncensored-ONNX",
  {
    device: "webgpu",
    dtype: "q4f16", // use "q4" for the Q4 variant
  },
);

const messages = [
  { role: "system", content: "You write concise English and Chinese image prompts." },
  { role: "user", content: "Create a cinematic prompt for a rainy neon street." },
];

const output = await generator(messages, {
  max_new_tokens: 256,
  temperature: 0.3,
  repetition_penalty: 1.05,
  streamer: new TextStreamer(generator.tokenizer, {
    skip_prompt: true,
    skip_special_tokens: true,
  }),
});

console.log(output[0].generated_text.at(-1).content);

The first browser load downloads roughly 1.7–1.9 GB. Use a progress indicator and let the browser cache the model for subsequent sessions. WebGPU support is required for the recommended Q4F16 path.

Files

config.json
generation_config.json
tokenizer.json
tokenizer_config.json
chat_template.jinja
onnx/model_q4.onnx
onnx/model_q4.onnx_data
onnx/model_q4f16.onnx
onnx/model_q4f16.onnx_data

The transformers.js_config section in config.json maps the external-data files and selects FP32 KV cache for Q4 or FP16 KV cache for Q4F16.

Provenance

• Source model: SC117/LFM2.5-2.6B-Uncensored
• Source revision: 578ad81f31061f16797ce9f5451835f59d1ad2d5
• Exporter: Liquid4All/onnx-export
• Exporter revision: 9a23ddd23035165f7414a5de3220a51e85780f64
• Export command: LiquidONNX LFM2 builder, block size 32, symmetric Q4; Q4F16 converted from the Q4 graph with FP16 runtime/cache tensors.

SHA-256

License

The source model is distributed under the LFM Open License 1.0. See LICENSE and the source model card for the applicable terms.