Minico M1 1105 ONNX

ONNX builds of Mincofficial/Minico-M1-1105, a 350M reasoning model with five effort levels (off, low, medium, high, max), for Transformers.js and ONNX Runtime. See the main repo for benchmarks and details.

Try it in your browser: 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, 433 tokens, ONNX Runtime 1.31 CPU). The full-precision build is below 100% because of small floating-point differences between runtimes. Use q8 unless the download size matters more than accuracy.

Run it with Transformers.js

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

const id = "Mincofficial/Minico-M1-1105-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: 2368,
  do_sample: true, temperature: 0.1, top_k: 50, repetition_penalty: 1.05,
  streamer: new TextStreamer(tokenizer, { skip_prompt: true }),
});

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 Minico M1 1105's weights put in. q8 and q4 quantize only the weights, in blocks of 32 (MatMulNBits, and GatherBlockQuantized for the embedding), in the same layout as Liquid AI's q4 build. Generated text was checked token for token against the PyTorch model.

Effort levels

The chat template also accepts aliases (none, minimal, xhigh, and others) and enable_thinking=False, which means off. Without a reasoning_effort, and for unknown values, it uses high (the default since 2026-11-08; it was medium). Recommended sampling: temperature 0.1–0.3, top_k 50, repetition penalty 1.05.

Pick one effort per conversation. Switching the effort, or turning thinking on or off, partway through a chat is not recommended: the model was trained on chats that keep one setting from start to finish. Start a new chat to change it.

License

Minico M1 1105 is a modified version of LiquidAI/LFM2.5-350M and is released under the LFM Open License v1.0 (see LICENSE). The weights were changed by fine-tuning (SFT, DPO, GRPO) and weight merging, as described above. 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.