Learn · Popular Open Source Models · Gemma & GLM Families
Google Gemma 2 (2B, 9B, 27B)
In plain English
Gemma 2 from Google DeepMind punches way above its weight. The 27B model outperforms much larger models while fitting completely onto a single 24GB graphics card! 💎
What is actually going on
Gemma 2 introduces sliding window local attention interleaved with global attention every 2 layers, logit soft-capping (capping attention logits at 50.0 and output logits at 30.0 to prevent training instability), and knowledge distillation from massive Gemini teacher models. Gemma 2 27B in Q4_K_M consumes ~16GB VRAM with outstanding benchmark performance.
Why it matters
Gemma 2 27B is one of the most efficient single-GPU models available, delivering top-tier performance on consumer RTX 3090/4090 and 36GB+ MacBooks.
Another way to think about it
A compact sports coupe with an engine tuned to beat heavyweight sedans on any track. 🏎️
Check yourself
What technique does Gemma 2 use to prevent activation logits from exploding?
- Logit soft-capping
- Zeroing out weights
- FP64 precision
- Random dropout
Show answer
Logit soft-capping — Logit soft-capping limits extreme values in attention matrices, improving numerical stability and output quality.