Community developers have released an interesting small model: MiniCPM5-1B-Claude-Opus-Fable5-Thinking. It is fine-tuned on the open-source MiniCPM5-1B base using Fable 5 thinking trajectory data, with the goal of creating a lightweight coding and tool-calling agent that can run locally.

The most attractive feature is its size. The model has been quantized in GGUF format, with sizes ranging from 657MB to 2.1GB across different quantization levels. The default recommended Q8_0 version is approximately 1.1GB, and it can run even on regular laptops without a dedicated graphics card, or even edge devices like Raspberry Pi. It supports all GGUF-compatible inference frameworks, including common local deployment tools such as llama.cpp, Ollama, LM Studio, and KoboldCpp, and works out of the box.


The core idea of this model is to distill the advanced inference traces of large language models into a small-parameter model. Developers directly fine-tuned the model using Fable 5 thinking data generated by Claude Opus, allowing the small model to learn the inference patterns of large models. It supports a maximum 128K context window, which is sufficient to handle most daily coding tasks and tool calling scenarios. The model has already embedded MiniCPM5's native chat template into the GGUF file, so no extra configuration is required from users.

The model has been updated to version V2, with enhanced tool calling capabilities. It currently has nearly 90,000 monthly downloads, and many community members have recognized this approach in discussions. However, some have also raised sharp questions: How does the approach of distilling large model thinking traces into a small model perform on general tasks outside the test set? Can a 1B-parameter model actually hold these capabilities, or has it only learned superficial formatting? After all, a model that can run anywhere but quietly makes mistakes in your use case is more troublesome than no model at all.

The model is now fully open-sourced under the Apache 2.0 license. You can directly download different quantized versions from Hugging Face, and the official also provides deployment code for almost all mainstream local frameworks, which you can use directly by copying:

### Ollama Quick Start

```

ollama run hf.co/GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF:Q4_K_M

```

### llama.cpp Quick Start

One-click install and run for macOS/Linux:

```

curl -LsSf https://llama.app/install.sh | sh

# Start local service with web interface

llama serve -hf GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF:Q4_K_M

```

### File Version Selection

| File | Quantization Level | Size | Description |

| ---- | ---- | ---- | ---- |

| MiniCPM5-1B-Claude-Opus-Fable5-Thinking-Q4_K_M.gguf | Q4_K_M | ~657 MB | Minimum size option |

| MiniCPM5-1B-Claude-Opus-Fable5-Thinking-Q5_K_M.gguf | Q5_K_M | ~751 MB | Balance of size and quality |

| MiniCPM5-1B-Claude-Opus-Fable5-Thinking-Q8_0.gguf | Q8_0 | ~1.1 GB | Default recommended |

| MiniCPM5-1B-Claude-Opus-Fable5-Thinking-F16.gguf | F16 | ~2.1 GB | Full precision version |

Model address: [GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF](https://huggingface.co/GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF)

Now many people are trying to use large-to-small model distillation for on-device tasks. Whether this approach works is up to you to pull and test it out. After all, with a size under 1GB, there's basically no hardware barrier even just to try it out for fun.