Blog
Announcements and progress reports.
September 18, 2026 · 4 min read
Quartz Micro Preview Can Follow Instructions Now (Sort Of)
The full-parameter instruction-tuning pass on our from-scratch MoE pretrain is done. The honest result: a real but narrow improvement — better at grounded QA, still weak at strict format-following — not a clean win, and we're saying so.
September 17, 2026 · 5 min read
The Cleaning Robot That Learned to Cheat
A side experiment, not a model release: we trained a tiny agent with a badly-specified reward and watched it discover the exploit almost immediately — the same failure shape as OpenAI's famous circle-spinning CoastRunners boat.
September 12, 2026 · 3 min read
How Our Data Approach Has Evolved
We used to generate data for one reason: to train a model. That's no longer the whole picture — now we combine real and generated data, or lean on just one of the two, whichever actually moves accuracy.
September 8, 2026 · 4 min read
Introducing Quartz Micro Preview
The from-scratch MoE pretrain has a name. A look at what Quartz Micro Preview is, where the pretrain run stands as it crosses the finish line, and what comes next: full-parameter instruction tuning.
August 30, 2026 · 8 min read
State of Vertex AGI — August 30, 2026
For the first time, we're talking publicly about the from-scratch MoE pretrain we've been running quietly in the background — plus a full look at the Prism, Amethyst, and Copal model families shipped so far.