Minico M2
Minico M2 is a 1B-class local text-generation model from Mincofficial. It is an independent fine-tune of MiniCPM5-1B for concise answers, visible reasoning, and practical tool-oriented conversations.
what it is
M2 keeps the MiniCPM5 chat format and tokenizer. The source format supports thinking and tool definitions, and the training mix contains ordinary conversations, visible reasoning traces, and tool-use examples. This card does not claim that M2 implements the exact effort-control interface of Minico M1.
This is a regular Minico M2 release. It is not an official ProCreations model and ProCreations did not endorse or publish this checkpoint.
how it learn
- base model: OpenBMB MiniCPM5-1B
- local base used for training: mlx-community/MiniCPM5-1B-4bit
- training method: MLX QLoRA on Apple Silicon
- trainable layers: 8 transformer layers
- LoRA rank: 8
- LoRA scale: 20
- learning rate: 2e-4
- maximum training length: 2,048 tokens
- training data: a balanced 4,000-example slice, 2,000 examples from each listed source
- split material: 726 validation examples and 726 test examples were retained from the prepared data
- sampling seed: 4242
- training seed: 42
- release checkpoint: the durable step-200 checkpoint, with validation loss 0.671 and training loss 0.957
The source data was normalized only where needed for MiniCPM5's chat template: serialized tool-call arguments were restored to JSON objects. No benchmark score is reported here, and the loss values above are training diagnostics rather than a comparison with another model.
use
For MLX-LM:
mlx_lm.generate \
--model Mincofficial/Minico-M2 \
--prompt "Explain why the sky looks blue in two short paragraphs." \
--max-tokens 256
The tokenizer's chat template should be used by applications that need system messages, thinking controls, or tools. Runtime support for those fields can vary by inference stack.
formats
- This repository contains the fused MLX model and tokenizer files.
- The Q5_K_M GGUF is published separately at Mincofficial/Minico-M2-GGUF.
- The GGUF filename is
Minico-M2-Q5_K_M.gguf.
On this Mac, the training path used MLX QLoRA because the core Unsloth CUDA training path is not an Apple MLX backend. The GGUF target was produced as F16 and quantized with llama.cpp's Q5_K_M quantizer, the same Q5_K_M quantization family exposed by Unsloth's GGUF workflow.
Minico M1 comparison
This comparison is factual and is not a standardized head-to-head benchmark.
| Minico M1 | Minico M2 | |
|---|---|---|
| base | LiquidAI/LFM2.5-350M | openbmb/MiniCPM5-1B |
| scale | 350M-class | 1.08B parameters |
| fine-tuning direction | QyrouNnet-AI/exp-reasoning-effort-control, as described by the M1 card | ProCreations dataset mix, MLX QLoRA, 8-layer adapter |
| reasoning interface | visible <think>...</think> blocks and effort presets |
MiniCPM5 chat template with thinking/tool fields; no claim of the M1 effort API |
| published artifacts | M1 repo contains a float16 dequantized export derived from its retained Q8_0 artifact | fused MLX model here and a separate Q5_K_M GGUF repo |
| release name | Minico-M1-Preview is the historical repository name |
Minico-M2 |
known rough edges
This checkpoint was trained on a balanced subset rather than the full source collections because local Apple Silicon throughput made a full long-context pass impractical. Long examples are capped at 2,048 training tokens. The model may be repetitive, over-explain, or emit visible reasoning when a downstream application would prefer a final answer only. Validate outputs before using it for consequential work.
credits and inspiration
A lot of credit goes to ProCreations. The dataset mix and the decision to explore a compact model with visible reasoning and tool-use behavior were inspired by his Grug series. His public work helped establish the direction for this experiment and made the source datasets available for others to learn from.
Please see and credit the originals:
Minico M2 is not a Grug rebrand, an official ProCreations release, or a claim of shared authorship. It is an independent Mincofficial model that acknowledges the inspiration and the open data sources directly.
Additional credit goes to OpenBMB for MiniCPM5-1B and to the MLX community conversion used to make local Apple Silicon training practical. The base model's Apache-2.0 terms apply; review the original dataset cards and licenses before redistributing derivatives.
license
Apache-2.0 for this release, subject to the terms and attribution requirements of the base model and the source datasets.
- Downloads last month
- 132
4-bit
Model tree for Mincofficial/Minico-M2
Base model
openbmb/MiniCPM5-1B