About
Open-source distillation, built in public
We distill frontier reasoning models into much smaller open models and publish the datasets to help the community train, evaluate, and deploy locally.
Our Mission
Frontier AI models from Anthropic, OpenAI, and Google are incredibly capable but require API access and can be expensive to use at scale. We believe the open-source community deserves access to similar reasoning capabilities.
Our approach is simple: we create high-quality datasets by querying frontier models with diverse prompts, then fine-tune open-source base models on these reasoning traces. The result is smaller, locally-runnable models that capture much of the original model's reasoning style.
All our work is open source. We use Unsloth for efficient fine-tuning and release models in GGUF format for easy local deployment.
How We Work
Dataset Creation
We curate diverse prompts (coding, math, science) and query frontier models with high reasoning effort to capture detailed traces.
Fine-tuning
We fine-tune open-source base models (mostly Qwen3 variants) on our reasoning datasets using Unsloth for faster, cheaper training.
Quantization & Release
We export GGUF with multiple quant levels (Q3/Q4/Q6/Q8) for llama.cpp so you can run locally on consumer hardware.
Support Our Work
We're college students funding this research ourselves. Creating high-quality datasets from frontier models isn't cheap - our Claude Opus dataset alone cost over $52 to generate. If you find our models useful, please consider supporting us.