Google

Gemma 4

Open source4 variants

gemma-4-e2b

instruct2.3B effective / 5.1B with embeddings

Google's Gemma 4 E2B — ultra-compact 2.3B effective parameter multimodal MoE for on-device deployment with full audio, image, and video support.

131K tokensFree / Open weightsimageaudiovideoMoEApache 2.0

LiveCodeBench

gemma-4-e2b

29.1%

gemma-4-31b

80.0%

gemma-4-26b-a4b

77.1%

gemma-4-e4b

44.0%

LiveCodeBench v6

instruct26B total / 4B active

Google's Gemma 4 26B MoE with 4B active parameters — frontier multimodal intelligence at efficient compute cost with 256K context.

262K tokensFree / Open weights

instruct31B

Google's 31B Gemma 4 model — top of the Gemma 4 family with strong reasoning, coding, vision, and 128K context.

262K tokensFree / Open weightsimage

instruct4.5B effective / 8B with embeddings

Google's Gemma 4 E4B — 4.5B effective parameter multimodal MoE supporting text, image, audio, and video with 128K context.

131K tokensFree / Open weights