In September 2026 I posted a series of LinkedIn updates highlighting a toy‑scale expert‑router model, new Gemini 3.x live models, the Jev/Needle semantic‑search Chrome extension, and Dream‑RSI research that drastically reduced agent calls.
The expert‑router model has a total of 540M parameters.
At any moment about 109M of those parameters are active on‑card.
It kept 99.84% of its progress on the 7 subjects it never read.
Only 54 of its 136 experts changed during the entire chess read.
The shared core learns 10x slower than the experts.
Gemini 3.8 Live and Live Extended Thinking launched in private preview for Gemini Enterprise.
Voice input reached GA in Gemini Enterprise.
Jev takes your query and the page's passages, scores every passage, and picks the strongest sentence in each, all in a single pass.
There are no embeddings and no vector index anywhere in the stack.
Dream‑RSI achieved the result in 317 agent calls, whereas the previous best system needed 51,200.