Gemma 3 (27B)

Gemma 3 is Google DeepMind's third-generation open model family, released in 1B / 4B / 12B / 27B sizes. The 27B flagship is multimodal (text + image + short video input), supports 128K tokens of context, 140+ languages, function calling and structured output — and fits on a single GPU/TPU. Weights are open under the custom Gemma Terms of Use (verified via the license:gemma tag on the HF repo; not an OSI-approved license, so check terms before commercial redistribution). Google positions it as "the most capable model you can run on a single GPU or TPU".

Independent (human preference): LMArena Elo 1338 for Gemma 3 27B — ahead of Llama-405B, DeepSeek-V3 and o3-mini, just below DeepSeek-R1 (1363), widely reported by independent tech press (ZDNet, Beebom). Provider-reported: Google's tech report claims the 27B is comparable to Gemini-1.5-Pro across benchmarks; MMLU-Pro 67.5%, GPQA Diamond 42.4% (press-reported from Google's materials). Google's "98% of DeepSeek-R1 on one GPU vs 32 GPUs" efficiency framing is Google's own estimate — treat as provider-claimed.