Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Minico M1

A 350M fine-tune of Liquid's LFM2.5 with a five-step thinking dial; tidy and tiny, but thinking barely helps it.

Reviewed by ReviewRamble editors · Updated

Visit Minico M1 Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.3/5

Minico M1 is a careful hobbyist fine-tune that makes a 350M model behave, but its headline feature, the thinking dial, does not make it more accurate.

What is Minico M1?

Minico M1 1105 is a 350-million-parameter reasoning model from Mincofficial, a solo developer, fine-tuned from Liquid AI's LFM2.5-350M and published on Hugging Face on 5 November 2026. Its distinctive feature is a reasoning-effort setting with five levels (off, low, medium, high and max) that changes how long the model thinks before it answers, from no reasoning at all to a median of 78 reasoning tokens. It also makes tool calls in LFM's format.

It replaces Minico M1 Preview, which the author says looped, repeated its reasoning in the answer and ignored the effort setting. Training ran in four stages (two rounds of supervised fine-tuning, DPO and GRPO with verifiable rewards) plus a weight merge, all on a 32GB MacBook Air with MLX. Builds come in safetensors, GGUF, MLX and ONNX, and a browser demo runs it on your own GPU through WebGPU.

Is Minico M1 worth it?

Minico M1 is the best-behaved 350M chat model in its author's tests, and the margin over its own base model is thin. On the five-benchmark harness in the model card it averages 49.8, against 47.0 for the untouched LFM2.5-350M, 43.6 for IBM's Granite 4.0 350M and 24.2 for Google's Gemma 3 270M. The author notes that differences under about 3 points are within run-to-run noise, which covers most of the gap to the base model.

The real gains are in behaviour rather than knowledge. Against the base model it moves multi-turn answers that address the latest message from 91 to 99 out of 100, and correct tool calls on the author's test from 54.8% to 100%. Loop rate, the share of 80 prompts where output repeats until the token limit, is 0%, against 18.8% for the Preview and for Gemma 3 270M. At about 229MB as a Q4_K_M GGUF, the card reports 371 tokens a second on a fanless MacBook Air.

The effort dial is honest but mostly cosmetic for now. On GSM8K grade-school maths the model scores 39.8 with thinking off and 37 at its best with thinking on, and the card says so plainly. Thinking does fix some traps: asked "What is 5 + 5 x 5?" 20 times, it was right 7 times with thinking off and 20 times at every other level.

The catch: this is a 350M model, so facts and harder maths go wrong, and only English was tested. Every number comes from one person's harness with near-greedy sampling, not from independent evaluation. With tools available it skipped an unneeded call only 75% of the time, and the LFM Open License withdraws commercial rights from any company with $10 million or more in annual revenue.

Our take: as a command parser for a voice assistant or a browser extension that must run offline, Minico M1 is a cleaner starting point than the stock LFM2.5-350M; as a small reasoning model, it shows that a dial on a 350M model changes how much it says, not how much it knows.

Where it earns its place

  • A developer building an offline voice assistant for a smart-home hub

    Turn spoken commands into function calls such as set_volume(level=30)

    Pass the device functions through `tools` in the chat template and run the 229MB Q4_K_M GGUF with llama.cpp. Keep effort off for speed and check that it does not call a tool when the user only chats, which it got wrong 25% of the time in testing.

  • A front-end engineer adding private text help to a browser extension

    Rewrite or shorten text in the page without sending it to a server

    The ONNX build runs through WebGPU in the browser, as the author's demo does, so nothing leaves the user's machine. There is no hosting bill; the cost is the download and the user's GPU.

  • A researcher studying small reasoning models

    Compare how reasoning length affects accuracy at 350M parameters

    Set reasoning_effort from off to max and log median reasoning tokens (40 at low, 78 at max per the card). The published training recipe, SFT, DPO, GRPO and ExPO merge, is reproducible on a 32GB Mac with MLX.

Before you commit

Setup
Load Mincofficial/Minico-M1-1105 with Transformers, or run `llama-server -hf Mincofficial/Minico-M1-1105-GGUF:Q8_0 --jinja` and pass reasoning_effort through chat template kwargs; an MLX build covers Apple silicon.
Integrations
Transformers, llama.cpp, MLX and ONNX Runtime (WebGPU in the browser); tool calls use LFM-style <|tool_call_start|> tags rather than OpenAI JSON.
Data and privacy
Weights run locally; the browser demo runs on your GPU and the author says nothing leaves your computer. Training data was distilled from Liquid's LFM2.5-1.2B-Thinking and LFM2.5-1.2B-Instruct.
Cost at scale
No usage fee. Commercial use is licensed only below $10 million in annual revenue; above that, the licence grants no commercial rights and separate terms are not published.
Limits
32K-token context from the base model; recommended output caps from 640 tokens (low) to 4,608 (max). The author advises one effort level per conversation, English only, and no medical, legal or financial use.

What’s good

  • Five reasoning-effort levels, from off to a median of 78 thinking tokens
  • 100% correct tool calls on the author's test, against 54.8% for the base model
  • 0% loop rate across 80 prompts
  • About 229MB as Q4_K_M; 371 tokens a second on a fanless MacBook Air
  • GGUF, MLX and ONNX builds plus an in-browser WebGPU demo

What’s not

  • Thinking does not raise GSM8K accuracy: 37 at best against 39.8 with it off
  • Average lead over its base model (49.8 vs 47.0) is close to the stated noise
  • All benchmarks are the author's own harness; English only
  • Commercial use barred for companies with $10 million or more in revenue

Best for

Developers building offline assistants, browser extensions or on-device command parsers who need a sub-400M model that follows instructions and calls tools reliably.

Minico M1 alternatives

Minico M1: common questions

Is Minico M1 worth it?
Minico M1 is a careful hobbyist fine-tune that makes a 350M model behave, but its headline feature, the thinking dial, does not make it more accurate. We rate it 3.3 out of 5.
How much does Minico M1 cost?
Free weights under the LFM Open License v1.0. Commercial use is licensed only for organisations with annual revenue under $10 million; non-profit research is exempt.
What are the downsides of Minico M1?
Thinking does not raise GSM8K accuracy: 37 at best against 39.8 with it off; average lead over its base model (49.8 vs 47.0) is close to the stated noise; all benchmarks are the author's own harness; English only; commercial use barred for companies with $10 million or more in revenue.
Who is Minico M1 best for?
Developers building offline assistants, browser extensions or on-device command parsers who need a sub-400M model that follows instructions and calls tools reliably.
How much does Minico M1 cost for a team or in production?
No usage fee. Commercial use is licensed only below $10 million in annual revenue; above that, the licence grants no commercial rights and separate terms are not published.
Is Minico M1 safe for company data?
Weights run locally; the browser demo runs on your GPU and the author says nothing leaves your computer. Training data was distilled from Liquid's LFM2.5-1.2B-Thinking and LFM2.5-1.2B-Instruct.
What are the best Minico M1 alternatives?
The closest alternatives we have reviewed are LFM3, Granite 4.2, SmolLM3-3B and Gemma 4.

Answers checked Nov 21, 2026.