DeepSeek-R1 is a strong performing, open source reasoning model with several distilled versions that can be run locally. Here, we walk through the DeepSeek-R1 paper to review the details of the training methodology, download the 14b distilled model via Ollama, test generation / JSON-mode, and then test it in a fully local "deep research" assistant that performs web-research / summarization w/ an iterative reflection step to improve its results.
Video notes:
https://mirror-feeling-d80.notion.site/DeepSeek-R1-182808527b17801585dadb84f7c66cd9?pvs=4
Code repo:
https://github.com/langchain-ai/ollama-deep-researcher
Related video on reasoning models:
https://www.youtube.com/watch?v=f0RbwrBcFmc
Prior build-from-scratch video for an earlier version of this research assistant:
https://www.youtube.com/watch?v=XGuTzHoqlj8