Gemma 4 31B tops GPQA Diamond

The interesting part here is not just the score, but the implied efficiency curve: if the benchmark holds up, Gemma 4 is squeezing near-frontier reasoning into a much more deployable footprint.

• –Google’s official launch says the 31B dense model fits on a single 80GB H100, which makes this feel less like lab bragging and more like something teams can actually run.
• –The Reddit post’s token-efficiency claim is the real differentiator: similar benchmark performance with fewer output tokens suggests lower inference cost per useful answer.
• –Gemma 4’s 256K context, multimodal input, and native function-calling make it more than a chat model; it’s clearly aimed at agentic workflows and local developer tooling.
• –The caution flag is provenance: this specific Qwen comparison is a community benchmark claim, not an official Google benchmark, so it should be treated as promising but not definitive.
• –Still, Apache 2.0 plus open weights means adoption friction is low, which is exactly what the open-model ecosystem needs right now.