Hugging Face, 350M parameters

Minico M1

A small reasoning model that runs on your own computer. You choose how hard it thinks, from off to max. After the Preview, this one actually works.

Try it in your browser Get it on Hugging Face All projects

Parameters350MAbout 229 MB as Q4_K_M.

Effort levels5Off, low, medium, high and max. Off answers right away.

Loop rate0%The Preview got stuck repeating itself on 18.8% of prompts.

What it does

It thinks first, then answers once.

M1 writes its reasoning between think tags, closes them, and gives one answer, without repeating the reasoning back to you. The effort setting decides how long it thinks. It is built for small everyday jobs: quick math, rewriting, short questions and simple tool calls, like the commands a voice assistant would run.

Real output

minico-m1

> What is 5 + 5 x 5?effort: mediumOkay, let's see. The user asked for 5 + 5 times 5. Hmm, first, I need to follow the order of operations. The numbers are in parentheses, so maybe they want me to evaluate the expression inside first. Then add that result to 5.To solve the expression 5 + 5 × 5, follow these steps: 1. **Evaluate the multiplication**: 5 × 5 = 25. 2. **Add the result to 5**: 5 + 25 = 30. Thus, the final answer is 30.

01
Pick an effort
Pass reasoning_effort: off, low, medium, high or max. Medium is the default.
02
It reasons
Higher efforts think longer: about 40 tokens at low, 78 at max.
03
It answers once
The think block closes and one clean answer follows. No second copy of the reasoning.
04
It calls tools
Give it functions and it calls them: 100% correct calls on our tool test.

Versus the Preview

Same questions, same settings. The Preview broke as soon as it thought; M1 does not.

GSM8K, thinking on37The Preview's best with thinking on was 6.5.

IFEval72.5Following instructions. The Preview got 50.3.

Multi-turn99Answers the latest message, not an old one. The Preview got 66.

Leads models its size

Minico M1 beats every other sub-400M model here on at least 4 of 5 benchmarks, and has the highest average, Qwen2.5 0.5B included. Scores out of 100, on our held-out test questions. Every model ran on the same harness with its own chat template.

Model	Size	GSM8K	MATH-500	IFEval	MMLU-Pro	Multi-turn	Average	Loops
Minico M1 (off)	350M	39.8	19	72.5	18.8	99	49.8	0%
Minico M1 Preview	350M	25.8	15.5	50.3	11.2	66	33.8	18.8%
Gemma 3 270M	270M	17.2	6	26.1	11.5	60	24.2	18.8%
Granite 4.0 350M	350M	41.5	17	48.8	12.6	98	43.6	1.2%
LFM2 350M	350M	26.8	22	55.3	17.1	71	38.4	0%
LFM2.5 350M (base)	350M	35	19	71.7	18.4	91	47	0%
SmolLM2 360M	360M	11.2	0	32.4	10.8	44	19.7	2.5%
Qwen2.5 0.5B	494M	47.8	36	28.8	14.6	81	41.6	7.5%

GSM8K: 400 questions. MATH-500: 200. IFEval: 541 prompts, strict. MMLU-Pro: 980 questions. Multi-turn: 100 chats. Average: mean of those five. Loops: share of 80 prompts where the output repeats until the limit. Temperature 0.1. M1 shown with thinking off, the same way the other models answer.

What it is still bad at

Hard math

Thinking makes the answer clearer, not much more correct. On GSM8K, off scores about the same as high.

Knowing when not to call a tool

With tools available, it skipped the tool when it should only 75% of the time.

Facts

It is 350M parameters. Check anything that matters, and do not use it for medical, legal or money advice.

Run it

$ llama-server -hf Mincofficial/Minico-M1-1105-GGUF:Q8_0 --jinja

You need

Any computer from the last few years. 371 tokens a second on a MacBook Air (Q4_K_M, llama.cpp). Or skip the install and try it in your browser.

Efforts

In llama.cpp, add --chat-template-kwargs with a reasoning_effort of off, low, medium, high or max. In Transformers, pass reasoning_effort to apply_chat_template. Keep one effort per chat: switching midway isn't recommended.

License

Built on LFM2.5-350M by Liquid AI, under the LFM Open License v1.0. Free for research and for companies under $10M a year in revenue.