Curiosity · AI Model
Gemma 3 27B
Gemma 3 27B is Google DeepMind's March 2025 open-weights flagship — a 27B dense transformer derived from Gemini research with native vision input, 128K context, and multilingual pretraining. Released under the Gemma license, it's the strongest single-GPU open model Google has shipped to date.
Model specs
- Vendor
- Family
- Gemma 3
- Released
- 2025-03
- Context window
- 128,000 tokens
- Modalities
- text, vision
- Input price
- $0.1/M tok
- Output price
- $0.2/M tok
- Pricing as of
- 2026-04-20
Strengths
- Open weights under Gemma license — commercial use permitted
- Native multimodality — text and vision from pretraining
- Single-GPU deployable at FP16 on H100 80GB
- Strong multilingual coverage — 140+ languages in pretraining
Limitations
- Gemma license includes responsible-use restrictions
- Trails Llama 4 Maverick and DeepSeek V3 on top-end reasoning
- Vision quality below Gemini 2.5 Flash — good but not SOTA
- Tool-use and agentic fine-tuning less polished than closed Gemini
Use cases
- Multimodal assistants combining text and vision on-prem
- Document RAG with mixed text and screenshot content
- Research baselines for Gemini-adjacent open models
- Fine-tuning base for enterprise vision-language workloads
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MMLU-Pro | ≈67% | 2025-03 |
| HumanEval | ≈85% | 2025-03 |
| LMArena ELO | ≈1338 | 2025-03 |
Frequently asked questions
What is Gemma 3 27B?
Google DeepMind's March 2025 open-weights flagship — a 27B multimodal dense transformer derived from Gemini research. Supports text and vision input with a 128K context window.
How does Gemma 3 compare to Gemini 2.5?
Gemini 2.5 Pro remains Google's closed flagship and is meaningfully stronger on agentic and long-context tasks. Gemma 3 27B is the open, self-host-friendly alternative at a much smaller footprint.
Can Gemma 3 27B run on a single GPU?
Yes — at FP16 it fits on an H100 80GB, and quantized to 4-bit it runs on a 24GB consumer GPU with moderate performance hits.
Sources
- Google — Gemma 3 blog — accessed 2026-04-20
- Hugging Face — google/gemma-3-27b-it — accessed 2026-04-20