Run DeepSeek-R1-0528 Dynamic 1-bit GGUFs
Blog post from Unsloth
DeepSeek-R1-0528 is an advanced open-source reasoning model developed by DeepSeek, competing with top models like OpenAI's GPT-4.5 and Google's Gemini 2.5 Pro. The model has been optimized through quantization to reduce its size from 720GB to 185GB, making it more accessible for diverse computing environments. Users can leverage Unsloth's 1.78-bit Dynamic 2.0 GGUFs for running the model on various inference frameworks, with specific settings recommended for optimal performance, such as a temperature of 0.6 and top_p of 0.95. The model is compatible with setups having at least 20GB RAM and can achieve higher throughput with more memory or a GPU, although it is technically possible to run it without a GPU by utilizing Apple's unified memory chips. Comprehensive guides and community support are available for users to effectively implement and test the model in applications, including gaming simulations, using Python and other computational libraries.