deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B · Hugging Face

DeepSeek has introduced first-generation reasoning models that validate large-scale reinforcement learning as a viable method for developing complex thought processes without initial supervised fine-tuning. This approach allows models to naturally discover powerful reasoning behaviors, such as self-verification and long chain-of-thought generation. By combining this raw reinforcement learning with targeted fine-tuning, the resulting models achieve performance parity with leading proprietary alternatives in mathematics, coding, and logical reasoning tasks. A critical finding from this research is the efficacy of distilling these advanced reasoning patterns into significantly smaller, dense models. These distilled versions retain high performance across various benchmarks while being far more accessible for local deployment. This democratization of high-level reasoning capabilities means that researchers and developers can now utilize state-of-the-art problem-solving skills without the immense computational resources required by massive foundation models. This release is highly relevant to the open_data community because it fully open-sources both the base reasoning models and the distilled variants under permissive licenses. By providing the underlying reasoning data and model weights, DeepSeek enables the community to build upon these capabilities, fostering transparency and accelerating innovation. This contribution supports a more open ecosystem where advanced AI reasoning tools are accessible for modification, commercial use, and further research, rather than being confined to closed APIs.

Source: huggingface.co
Published on 2025-01-22

Creció actividad económica 1.1% en 2024, estima Inegi »

Open Data

Related news