Minico M1 1105

The SOTA of 350M models. A reasoning model you can run on a laptop, and you choose how hard it thinks.

Minico M1 1105 is the full release of Minico M1, fine-tuned from LiquidAI/LFM2.5-350M. It fixes what was broken in Minico M1 Preview: the effort levels now work, the reasoning is no longer repeated in the answer, and it no longer loops.

Try it in your browser (runs on your GPU with WebGPU, nothing leaves your computer)

GGUF: Mincofficial/Minico-M1-1105-GGUF · MLX: Mincofficial/Minico-M1-1105-MLX · ONNX: Mincofficial/Minico-M1-1105-ONNX

Highlights

• Reasoning with five effort levels: off, low, medium, high, max. One of the first sub-400M reasoning models, and the only one we know of with a five-step effort dial.
• Thinks once, answers once: reasoning closes with </think> and one answer follows. Loop rate 0% (Preview: 18.8%).
• Leads the sub-400M models we tested: beats each of them on at least 4 of 5 benchmarks, with the highest average of every model we ran, Qwen2.5 0.5B included (49.8; next best LFM2.5 350M base at 47.0). Against Granite 4.0 350M: IFEval 72.5 vs 48.8, MMLU-Pro 18.8 vs 12.6, MATH-500 19 vs 17, multi-turn 99 vs 98; Granite is ahead on GSM8K (41.5 vs 39.8). Same harness, same questions; table below.
• Above Qwen2.5 0.5B, a model 41% bigger, on IFEval, MMLU-Pro, Multi-turn (IFEval 72.5 vs 28.8; MMLU-Pro 18.8 vs 14.6; Multi-turn 99 vs 81).
• Tool calling: 100% correct calls on our function-calling test (base LFM2.5-350M: 54.8%).
• Runs anywhere: ~229 MB as Q4_K_M, 371 tokens/s on a fanless MacBook Air with llama.cpp. GGUF and MLX builds included.

Benchmarks

All models ran on the same harness, on the same held-out test questions, with their own chat templates and the same sampling (temperature 0.1, top_k 50, repetition penalty 1.05). Minico M1 1105 is shown with thinking off, the way the other models answer. Best sub-400M score in bold. Average = mean of the five scored benchmarks (loop rate not included). Qwen2.5 0.5B is included as a larger reference. Every model used this same near-greedy sampling rather than its own recommended settings, so official numbers can differ: Qwen2.5 0.5B's IFEval here matches its published score, while Gemma 3 270M scores lower here than in Google's report.

GSM8K: 400 test questions. MATH-500: 200, scored only when the answer is in \boxed{} as the prompt asks (SmolLM2 rarely does this, which is most of its 0). IFEval: 541 prompts, prompt-level strict. MMLU-Pro: 980 questions (70 per subject). Multi-turn: 100 chats where the answer must address the latest message. Loop rate: share of 80 mixed prompts where the output repeats until the token limit. Differences under about 3 points are within run-to-run noise.

Versus the Preview

By effort

Honest note: on GSM8K, thinking makes the answer easier to follow but not more often right; off scores about the same as high. The effort levels change how long it thinks (median tokens above), and every level closes its reasoning and answers.

Example: "What is 5 + 5 x 5?"

Asked 20 times at each effort (temperature 0.1), correct answers out of 20. With thinking on, the Preview said 50 in 80 of 80 tries.

Thinking is what gets this one right: with off, Minico M1 1105 added first and said 35 in 12 of 20 tries. Written as 5 + 5 × 5, it answered 30 in 99 of 100 tries across all five efforts. One run of each effort:

Quickstart

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Mincofficial/Minico-M1-1105"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto")

messages = [{"role": "user", "content": "A shirt costs $25 and is 20% off. How much do I pay?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, reasoning_effort="high",
                                 return_tensors="pt", return_dict=True)
out = model.generate(**inputs, max_new_tokens=2368, do_sample=True, temperature=0.3, top_k=50, repetition_penalty=1.05)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False))

llama.cpp

llama-server -hf Mincofficial/Minico-M1-1105-GGUF:Q8_0 --jinja --chat-template-kwargs '{"reasoning_effort":"high"}'

MLX (Apple silicon)

mlx_lm.generate --model Mincofficial/Minico-M1-1105-MLX --prompt "What is 5 + 5 x 5?" --chat-template-config '{"reasoning_effort":"low"}'

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.

Tool calling

Pass tools with tools=[...] in apply_chat_template. The model answers with LFM-style calls: <|tool_call_start|>[set_volume(level=30)]<|tool_call_end|>. Return results with the tool role.

Training

1. SFT 1: reasoning traces at every effort level, distilled from LiquidAI/LFM2.5-1.2B-Thinking (thinking efforts) and LFM2.5-1.2B-Instruct (off), with length budgets per effort.
2. SFT 2: on-policy rejection sampling (correct, closed, non-looping answers only), verified instruction-following data, greetings and identity.
3. DPO: preferred vs rejected pairs for loops, reasoning repeated in the answer, and identity.
4. GRPO: verifiable rewards for math, instruction following, tool calls and multi-turn; penalties for loops, unclosed <think>, reasoning that contradicts the answer, calling tools when none are needed, and claiming another identity; effort length floors so higher efforts think longer.
5. Merge: extrapolation (ExPO) from the GRPO checkpoints, kept only because it scored higher on a separate dev split.

All training ran locally on a 32 GB MacBook Air with MLX.

Limitations

• It is a 350M model. It gets facts and harder math wrong. Do not use it for medical, legal or financial advice.
• With tools available, it sometimes calls a tool when it should just answer (it skipped the tool correctly 75% of the time on our test).
• Thinking does not raise GSM8K accuracy over off yet (37 at best vs 39.8 off).
• English only was tested.

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.