Thread · 3 stories · Jul 14 – Aug 26
Moonshot's 2.8T-parameter open-weight model, from announcement to Hugging Face to the arithmetic of whether anyone can afford to serve it.
Jump to timeline ↓Moonshot AI's K3 succeeded K2 as the largest open-weight reasoning model of the summer, hitting the Hacker News front page within hours of its weights going up.
The thread ends where open weights usually do: with someone working out whether anyone can actually afford to serve it, at 159 tokens per GPU-second on rented B200s.
