Gemma 4 vs Qwen3.8-27B: which open model should you run on one GPU?

Gemma 4 from Google and Qwen3.8-27B from Alibaba are Apache 2.0 models that fit one GPU. Compare sizes, input types, memory needs and where each fits.

  • Updated

Verdict. Both use Apache 2.0, read images and run on a single GPU. Gemma 4 is a family of five sizes, from the phone-sized E2B (about 3.5 GB at 4-bit by our estimate) to a 31B dense flagship, and the smaller sizes accept audio. Qwen3.8-27B is one 27.8B dense model that suits a single 24 to 48 GB GPU or a Mac with enough memory. The quality claims come from Google and Alibaba themselves, so try both on your own prompts.

Specifications

Specifications Gemma 4Qwen3.8-27B
DeveloperGoogle DeepMindAlibaba Qwen
ReleasedApril 2, 2026August 14, 2026
Total parameters30.7B (31B)27.8B (27B)
Active parametersAllAll
ArchitectureDenseDense
Context window256K tokens256K tokens
InputText, Image, Audio, VideoText, Image, Video
License Apache 2.0 (Permissive) Apache 2.0 (Permissive)
Memory for the smallest variant (4-bit)3.52 GB (E2B)19 GB (27B)

Performance claims come from the developer and have not been independently verified.

Pick Gemma 4 if

  • You need a model for small hardware: E2B and E4B are meant for on-device and offline apps.
  • You want one family with a size for every machine, from E2B up to a 31B dense model and a 26B mixture of experts that activates only 3.8B parameters.
  • You need audio input, which the E2B, E4B and 12B models accept.
  • You need answers in many languages: Google lists more than 35.

Pick Qwen3.8-27B if

  • You have one 24 to 48 GB GPU, or a Mac with enough memory, and want a single strong all-round model.
  • You want a ready-made FP8 build: Alibaba publishes a 31 GB FP8 checkpoint next to the 56 GB BF16 one.
  • You plan to fine-tune a dense model, which is simpler to train than a very large mixture of experts.

Gemma 4 · Qwen3.8-27B