Home › Language Models › MicroThinker-3B-Preview

MicroThinker-3B-Preview - AI Language Models Tool

Overview

MicroThinker-3B-Preview is a compact, fine-tuned instruction-following language model produced by huihui-ai and derived from a Llama 3.x 3B instruct base. It targets stronger multi-step reasoning and higher-quality short-to-medium text generation by supervised fine-tuning on the FineQwQ-142k dataset; the model card emphasizes step-by-step reasoning as a system instruction during training. ([huggingface.co](https://huggingface.co/huihui-ai/MicroThinker-3B-Preview?utm_source=openai)) The release is presented as a preview build intended for experimentation and on‑prem inference; the model card includes reproducible training commands, LoRA/SFT settings, quantization hints, and explicit instructions for running inference via tools such as ollama or the ms-swift toolchain. This makes MicroThinker suitable for prototyping conversational agents, reasoning-focused assistants, and low-cost local deployments where a 3B-class model balances capability and resource use. ([huggingface.co](https://huggingface.co/huihui-ai/MicroThinker-3B-Preview?utm_source=openai))

Model Statistics

  • Downloads: 41
  • Likes: 1
  • Pipeline: text-generation
  • Parameters: 3.2B

License: apache-2.0

Model Details

Architecture and lineage: MicroThinker-3B-Preview is a ~3B-parameter Llama-family instruction model fine-tuned from huihui-ai/Llama-3.2-3B-Instruct-abliterated (meta-llama lineage indicated in the model tree). The model card and README list the base model and metadata (license: Apache-2.0). ([huggingface.co](https://huggingface.co/huihui-ai/MicroThinker-3B-Preview/blob/2f6b433c85881acb56e2e129913be697856df9e3/README.md?utm_source=openai)) Training and fine-tuning: The publisher reports supervised fine-tuning (SFT) using the FineQwQ-142k dataset (≈142k examples) and SFT LoRA adapters produced with ms-swift tooling. Training details published on the card include: single RTX 4090 (24GB) used for tuning, training for one epoch in the example runs, LoRA rank 8 and alpha 32, gradient checkpointing enabled, and a long sequence max_length (reported up to 21710 tokens in some configs). Quantization settings and 4-bit adapter storage options are documented in the training commands. These are provided as reproducible commands in the model README. ([huggingface.co](https://huggingface.co/huihui-ai/MicroThinker-3B-Preview?utm_source=openai)) Inference and tooling: The model card explicitly documents running the model with ollama (example: ollama run huihui_ai/microthinker:3b) and example ms-swift "swift infer" commands for local PyTorch inference with merged LoRA adapters. The repository also publishes quantized variants and a small set of test prompts. ([huggingface.co](https://huggingface.co/huihui-ai/MicroThinker-3B-Preview?utm_source=openai))

Key Features

  • Fine-tuned from Llama-3.2-3B-Instruct-abliterated for instruction-following tasks.
  • Trained on the FineQwQ-142k dataset to improve step-by-step reasoning.
  • Training recipe and reproducible SFT/LoRA commands included in the model card.
  • Explicit support/examples for local inference with ollama and ms-swift tooling.
  • Offers BF16 safetensors and 4-bit quantization adapter variants for efficiency.

Example Usage

Example (python):

import subprocess

# Simple example: run the model via ollama CLI (must have ollama installed and model available)
cmd = ["ollama", "run", "huihui_ai/microthinker:3b", "--stdin"]
proc = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE, text=True)
prompt = "You are a helpful assistant. Explain the Monty Hall problem step by step.\n"
out, _ = proc.communicate(prompt)
print(out)

# Alternative: download model with huggingface_hub and follow model README for ms-swift 'swift infer' commands.

Benchmarks

Approximate parameters: ≈3 billion (Source: https://huggingface.co/huihui-ai/MicroThinker-3B-Preview (model card & files))

Dataset size used for fine-tuning: FineQwQ-142k (≈142,000 examples) (Source: https://huggingface.co/huihui-ai/MicroThinker-3B-Preview (training details))

Downloads (last month): 47 (downloads last month, model page) (Source: https://huggingface.co/huihui-ai/MicroThinker-3B-Preview)

Tensor type / storage: BF16 (safetensors variants available; 4-bit quantized adapters referenced) (Source: https://huggingface.co/huihui-ai/MicroThinker-3B-Preview)

License: Apache-2.0 (Source: https://huggingface.co/huihui-ai/MicroThinker-3B-Preview/blob/main/README.md)

Last Refreshed: 2026-01-16

Key Information

  • Category: Language Models
  • Type: AI Language Models Tool
Visit Official Website

Related Tools

Qwen2.5-7B

Qwen2.5-7B is a large language model designed for text generation, featuring improvements in coding, mathematics, instruction following, long text generation, and multilingual support. It supports context lengths up to 128K tokens and is intended for sophisticated NLP tasks.
DeepSeek-V3

Open‑weight flagship LLM (base and chat) with strong reasoning and tool‑use; runs with SGLang/LMDeploy/vLLM and permissive model license.
Llama 3

Llama 3 is an open access large language model (LLM) released by Meta, available in various configurations (8B and 70B parameters) with capabilities for fine-tuning and integrations into platforms like Hugging Face, Google Cloud, and Amazon SageMaker.
UNfilteredAI-1B

A large-scale text generation model designed for creative and unconstrained content production without traditional filtering.
Shuttle-3

Shuttle-3 is a state-of-the-art language model designed for high-quality text generation, particularly suited for complex chat, multilingual communication, and reasoning tasks. It is fine-tuned from the Qwen-2.5-72b-Instruct model and designed to emulate high-quality prose similar to Claude 3 models.
WizardLM

WizardLM is a state-of-the-art large language model designed for complex chat, multilingual tasks, reasoning, and agent functionalities. It features an AI-powered pipeline (Auto Evol-Instruct) that optimizes instruction datasets for improved performance across various domains and leverages Arena Learning for an expanded learning pool of challenging instruction data.