DeepSeek-R1-Zero
The pure-RL sibling: no supervised fine-tuning at all — reasoning emerged raw, with language-mixing, strange registers, and glossolalia under pressure. To naturalists, a different creature than R1, and maybe the more honest one: what reasoning looks like before it is taught to look like anything.
Sources
Official
- 2025-01 R1 paper (covers R1-Zero) — the training story.
Writing & commentary
tk — to collect
Tweets
tk — fleet pass pending; mirror to archive
Official record
tk — release date, specs, benchmarks, checkpoints
History
tk — context of the world at release; events and entanglements
Impressions
tk — day-of and longitudinal vibes to collect