GitHub radar
DeepSeek R1: Open Reasoning Model at o1 Level
DeepSeek's first-generation reasoning model matches OpenAI-o1 on math and coding benchmarks, carries an MIT license, and ships with distilled versions down to 7B for local use.
DeepSeek-R1 is DeepSeek's first-generation reasoning model, released in January 2025. According to the authors, it uses a mixture-of-experts architecture with 671 billion total parameters, of which only 37 billion are active per token. At the moment it is one of the most downloaded reasoning models on Hugging Face, with nearly 741,000 downloads in the past month. Regarding its capabilities, the model reasons through complex problems step by step, and according to DeepSeek it achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks. The MIT license permits both personal and commercial use without restrictions. DeepSeek also published distilled variants ranging from 1.5B to 70B parameters. In my view, the most practical starting point for local use is the 7B or 14B distilled version, which runs on a regular laptop.
Why a vibe-coder should care
This model is worth knowing about because it is the first open-weight reasoning model from a known vendor that approaches OpenAI-o1 quality. The full model is accessible via the DeepSeek API at platform.deepseek.com, and the chat at chat.deepseek.com with DeepThink mode enabled is free to try. If you want to run it locally without any API, the 7B distilled version needs about 4 GB in 4-bit format and fits on a laptop with 16 GB of RAM.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open https://huggingface.co/deepseek-ai/DeepSeek-R1, find the DeepSeek-R1 distilled 7B or 14B version and install it via Ollama — then show me how to run it and ask my first question
The full 671B model requires server-grade hardware and cannot run at home. The distilled 7B version needs about 4 GB in 4-bit format and runs on a laptop with 16 GB of RAM.
Open on Hugging Face▌ More finds