I'm speculating but Google might have figured out some training magic trick to balance out the information storage in model capacity. That or this flash model has huge number of parameters or something.
https://artificialanalysis.ai/evaluations/omniscience
Prepare to be amazed
Can someone explain how Gemini 3 pro/flash then do so well then in the overall Omniscience: Knowledge and Hallucination Benchmark?
> In the Hallucination Rate vs. AA-Omniscience Index chart, it’s not in the most desirable quadrant
This doesn't mean much. As long as Gemini 3 has a high hallucination rate (higher than at least 50% others), it's not going to be in the most desirable quadrant by definition.
For example, let's say a model answers 99 out of 100 questions correctly. The 1 wrong answer it produces is a hallucination (i.e. confidently wrong). This amazing model would have a 100% hallucination rate as defined here, and thus not be in the most desirable quadrant. But it should still have a very high Omniscience Index.
One hypothesis is that gemini 3 flash refuses to answer when unsuure less often than other models, but when sure is also more likely to be correct. This is consistent with it having the best accuracy score.
They have a similar chart that compares results across all their benchmarks vs. cost and 3 Flash is about half as expensive as 3 Pro there despite being four times cheaper per token.
That's what MoE is for. It might be that with their TPUs, they can afford lots of params, just so long as the activated subset for each token is small enough to maintain throughput.
More experts with a lower pertentage of active ones -> more sparsity.