sizzlebop/minimind-3-GGUF

MiniMind 3 GGUF

This repository provides GGUF quantizations for scikit-plots/minimind-3 (from the MiniMind series by Jingyao Gong).

MiniMind-3 is an ultra-compact language model (60M parameters) developed from scratch to demonstrate full-lifecycle LLM pretraining, supervised fine-tuning, and reasoning alignment at extreme efficiency. Despite its minuscule memory footprint (40 MB for Q4_K_M), it supports ChatML formatting, reasoning traces (<think>...</think>), and tool use.

All GGUF files were converted from the original safetensors weights using llama.cpp at native F16 precision, followed by standard k-quant quantization.

---

Available Files and Quantizations

File	Quant Type	Size	Description / Recommendation
minimind-3-F16.gguf	F16	122.16 MB	Full precision base conversion. Highest fidelity reference weights.
minimind-3-Q8_0.gguf	Q8_0	65.03 MB	Near-lossless 8-bit quantization. Recommended for best quality evaluation.
minimind-3-Q6_K.gguf	Q6_K	53.72 MB	High quality retention with minimal degradation.
minimind-3-Q5_K_M.gguf	Q5_K_M	46.07 MB	Balanced performance and memory footprint.
minimind-3-Q4_K_M.gguf	Q4_K_M	40.51 MB	Lightweight 4-bit quant. Recommended default for microcontroller and edge deployment.
minimind-3-Q3_K_M.gguf	Q3_K_M	32.94 MB	Ultra-compact footprint for tightly memory-constrained devices.
minimind-3-Q2_K.gguf	Q2_K	26.61 MB	Extreme compression (~26 MB).

---

Model Architecture Details

• Base Architecture: Qwen3ForCausalLM (qwen3)
• Parameters: ~60M (60,486,912)
• Layers: 8
• Hidden Size: 768
• Attention Heads: 8 query heads / 4 KV heads (GQA), head dimension 96
• Intermediate Size: 2432
• Context Length: 32,768 tokens
• Vocabulary Size: 6,400 (compact BPE tokenizer)
• RoPE Theta: 1,000,000.0

---

Prompt Format and Chat Template

MiniMind-3 utilizes ChatML format with optional <think> reasoning blocks:

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
What is 25 * 4?<|im_end|>
<|im_start|>assistant
<think>

---

Quickstart Guide

1. llama.cpp

Run interactive text generation with llama-cli:

llama-cli -m ./minimind-3-Q4_K_M.gguf \
    -p "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nHello! Who are you?<|im_end|>\n<|im_start|>assistant\n" \
    -n 256 \
    --temp 0.7 \
    --top-p 0.85

Launch a local inference server with llama-server:

llama-server -m ./minimind-3-Q4_K_M.gguf \
    --port 8080 \
    -c 4096

2. Ollama

Create a Modelfile:

FROM ./minimind-3-Q4_K_M.gguf
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>"""
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.7

Build and run:

ollama create minimind-3 -f Modelfile
ollama run minimind-3 "Explain what gravity is."

3. LM Studio

1. Copy the .gguf file to your LM Studio models folder.
2. Select minimind-3-Q4_K_M.gguf.
3. Choose ChatML preset (<|im_start|> / <|im_end|>).