Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-02 · high
Galileo's roughly 440-million-parameter Luna model scores production responses in milliseconds, catching quality drops before a weekly review would.
Why now: The push comes as more production LLM features ship without any live monitoring layer to catch drift between scheduled evals.
Small fine-tuned models can judge live LLM output instead of waiting for a weekly offline review, per a guide featuring Galileo's Luna model. Luna runs at roughly 440 million parameters, a fraction of a frontier model's size. It scores each response in milliseconds, at a fraction of what a frontier-model judge would cost.
Wire a model like Luna to a random sample of live traffic. A regression that would otherwise wait for the next scheduled review gets flagged the moment it ships.
The catch is that a judge model is only as good as the threshold behind it. Set the bar too loose and drift slides through anyway. Set it too tight and every noisy edge case pages someone. The guide frames Luna as a monitor wired to random sampling. Picking a fast model is the easy part; deciding what counts as a miss is the real work.
If you're running anything customer-facing on an LLM, a once-a-week eval doesn't catch what breaks on a Tuesday. A cheap judge model watching live traffic fixes that, but only for teams willing to tune the threshold instead of trusting the default.
Each link below shares sources, entities, or timing with this story.
Galileo released Luna / Shared entity: Luna / Shared topic / What happened next
Linked by a graph relationship (Galileo released Luna); both cover Luna; overlapping topics (frontier, model).
Galileo released Luna / Shared entity: Luna / What happened next / Tension
Linked by a graph relationship (Galileo released Luna); both cover Luna; picks up the Luna thread on 2026-07-10.
Galileo released Luna / Shared entity: Luna / Shared topic / What happened next
Linked by a graph relationship (Galileo released Luna); both cover Luna; overlapping topics (cost, model).
Linked by a graph relationship (Galileo released Luna); both cover Luna; overlapping topics (cost, model).
Galileo released Luna / Shared entity: Luna / Shared topic / Earlier coverage
Linked by a graph relationship (Galileo released Luna); both cover Luna; overlapping topics (frontier, model).
Linked by a graph relationship (Galileo released Luna); both cover Luna; overlapping topics (frontier, model).
Galileo released Luna / Shared entity: Luna / What happened next / Downstream implication
Linked by a graph relationship (Galileo released Luna); both cover Luna; picks up the Luna thread on 2026-07-19.
Galileo released Luna / Shared entity: Luna / What happened next
Linked by a graph relationship (Galileo released Luna); both cover Luna; picks up the Luna thread on 2026-08-07.