Minico M1 Preview ONNX

Deprecated. The Preview is the first Minico M1 checkpoint and is kept for comparison only. Use Minico M1 1105: it does better on our tests (see its model card) and can use tools.

ONNX builds of Mincofficial/Minico-M1-Preview, the first Minico M1 checkpoint: a 350M reasoning model with low, medium, high and max effort, for Transformers.js and ONNX Runtime. For the current model, use Minico M1 1105.

Try it in your browser: pick "Minico M1 Preview" (marked Deprecated) in the model menu. It runs the 8-bit build on your GPU with WebGPU, and nothing you type leaves your computer.

Agreement: share of tokens where the ONNX build's top choice matches the PyTorch model's greedy output (6 prompts, 347 tokens, ONNX Runtime 1.31 CPU).

Run it with Transformers.js

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

const id = "Mincofficial/Minico-M1-Preview-ONNX";
const tokenizer = await AutoTokenizer.from_pretrained(id);
const model = await AutoModelForCausalLM.from_pretrained(id, { dtype: "q8", device: "webgpu" });

const messages = [{ role: "user", content: "A shirt costs $25 and is 20% off. How much do I pay?" }];
const inputs = tokenizer.apply_chat_template(messages, {
  add_generation_prompt: true,
  return_dict: true,
  reasoning_effort: "high",
});
await model.generate({
  ...inputs,
  max_new_tokens: 1024,
  do_sample: true, temperature: 0.1, top_k: 50, repetition_penalty: 1.05,
  streamer: new TextStreamer(tokenizer, { skip_prompt: true }),
});

The prompt already opens the think block (<think>\n), so the output starts with the reasoning and then </think>. Pass enable_thinking: false to answer without thinking. The default effort is high. Its thoughts stay short at every level: 21 to 49 tokens on our test prompts.

WebGPU works in Chrome and Edge 113 or newer and in Safari 26. Without WebGPU, use device: "wasm" (much slower).

How it was made

The graph is Liquid AI's ONNX export of LFM2.5-350M (LiquidAI/LFM2.5-350M-ONNX) with the Preview's weights put in. Unlike later Minico models, the Preview has a separate output layer (lm_head), so the graph's logits read their own weight instead of the input embedding. q8 quantizes only the weights, in blocks of 32 (MatMulNBits, and GatherBlockQuantized for the embedding). Generated text was checked token for token against the PyTorch model.

The Preview's model.safetensors is itself a float16 export of a Q8_0 GGUF (see the main repo), which is why the 8-bit build tracks it so closely.

License

Minico M1 Preview is a modified version of LiquidAI/LFM2.5-350M and is released under the LFM Open License v1.0 (see LICENSE). Commercial use is allowed for organisations under the revenue threshold set in the license; read it before using the model commercially.

Built on LFM2.5-350M by Liquid AI. Attribution notices from the base model are kept in LICENSE.