Gemma 4 31B uses about 2.5x fewer tokens than Qwen3.5 27B but trails on agentic tasks, which helps pick between open models for local or cost-sensitive agents.
Key takeaways · AI-distilled
Gemma 4 comes in four sizes: 31B dense (39 on the Artificial Analysis Intelligence Index), 26B A4B mixture-of-expertsA model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.Full definition → (31), E4B (19) and E2B (15). The flagship is 29 points above Gemma 3 27B Instruct, which scored 10.
Compared with Gemma 3, all sizes add a reasoning mode and native image and video input (audio on E2B and E4B), context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → doubles to 256K on the larger models, and the license moves from Gemma Terms of Use to Apache 2.0.
The 31B trails Qwen3.5 27B by 3 points, which Artificial Analysis attributes mainly to weaker agentic results; it leads on SciCode (43% vs 40%) and TerminalBench Hard (36% vs 33%) and ties on GPQA Diamond and IFBench.
The small E4B and E2B models score better on AA-Omniscience (-20 and -24) than the 31B (-45). E2B has 2.3B active parameters and is designed for on-device deployment; at 4-bit quantizationShrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.Full definition → its model weights fit in under 3GB of RAM.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
