The content discusses deploying DeepSeek R1 671B locally with ollama, featuring models like DeepSeek-R1-UD-IQ1_M and DeepSeek-R1-Q4_K_M, with specific hardware requirements. The steps include downloading model files from HuggingFace, installing ollama, creating modelfiles, and running the model. Observations reveal the models’ performance on various prompts, showcasing the 4-bit version as “safer” in rejecting offensive or silly prompts. The conclusion suggests using Unsloth AI’s 1.73-bit version for lighter works due to slower generation speed with longer context length. The article invites readers to share their findings in the comments.
https://snowkylin.github.io/blogs/a-note-on-deepseek-r1.html