Friday 17 July 2026
An open three-trillion-parameter model you can download, a Victorian wildlife library brought back plate by plate, and an agent that reads a page's structure instead of a screenshot of it.
The open frontier hits three trillion.
Moonshot AI shipped Kimi K3, a mixture-of-experts model with 2.8 trillion total parameters, a one-million-token context, and native vision. It activates 16 of 896 experts per token, adds Kimi Delta Attention to hold long context cheaply, and lands as an open-weight release — full weights arrive 27 July, and the API already runs at $0.30 per million cache-hit tokens.
A Victorian wildlife library, restored plate by plate.
Nicholas Rougeux spent nearly a year rebuilding William Jardine's Naturalist's Library (1833–1843). Each of its hand-coloured plates was masked and colour-corrected in Photoshop — generative fill reassembled torn engravings — and 2.35 million OCR'd words were cleaned by hand. The result is a free web edition, a seven-pound printed book, and 27 posters.
- 40 volumes
- 1,347 plates
- 2.35M words
- 6,500+ links
- 4 divisions
The page describes itself to the agent.
Sentry rebuilt how its Seer agent perceives a screen. Instead of dumping an ASCII grid of the whole page, components self-register semantic context — titles, modes, filters, values — into a registry the agent queries on demand. On the most-visited issue pages, that context fell from ~2,100 tokens to ~300, with tool-call and satisfaction rates holding steady across 14,000+ eval runs.
That's Friday.
Three picks: an open frontier model you can run, a 190-year-old library brought back by hand, and an agent that trimmed its own context by 85%.