Local Thinking Model Released: OpenBMB’s MiniCPM5-1B Fine-Tuned on Claude Fable 5 Data

A community developer known as GnLOLot has unveiled a new local model named MiniCPM5-1B-Claude-Opus-Fable5-Thinking. This model, which is 1B in size, operates independently on local hardware and does not require any API keys or cloud connectivity. Built on the OpenBMB MiniCPM5-1B framework, it features a robust architecture with 1.08 billion parameters, 24 layers, and a context length of 131,072 tokens. The model is designed to facilitate reasoning through an enable_thinking toggle, allowing users to switch between thinking modes. The developer has fine-tuned the model using data from Fable 5, enhancing its coding abilities and instruction compliance. Unlike traditional distillation methods, this approach involves generating conversations with a teacher model and fine-tuning the smaller model based on those outputs. The resulting model learns to replicate the response style without inheriting the teacher's full capabilities. The model is available in various quantized formats, with the smallest version being approximately 657MB. Users can run it through platforms like llama.cpp and Ollama, following specific instructions provided in the repository.