Curiosity · AI Model
Gemma 3 4B
Gemma 3 4B is a 4-billion-parameter open-weights small language model from Google DeepMind, released March 2025. It is part of the third-generation Gemma family (1B / 4B / 12B / 27B), distilled from Gemini 2.0 research with substantial upgrades over Gemma 2: 128k context (up from 8k), native image input, tool-use friendly training, and multilingual coverage across 140+ languages. The 4B variant is the sweet spot for single-GPU or high-end laptop deployments.
Model specs
- Vendor
- Google DeepMind
- Family
- Gemma 3
- Released
- 2025-03
- Context window
- 128,000 tokens
- Modalities
- text, vision
Strengths
- Rare multimodal + multilingual combo at 4B
- 128k context in a tiny footprint
- Permissive Gemma licence
- Distilled from Gemini 2.0 research
Limitations
- Below 12B Gemma-3 and 70B-class models on deep reasoning
- Vision is image-only (no video)
- Licence has usage restrictions akin to Llama
- Fine-tune recipes still maturing on day-1
Use cases
- Multimodal agents on a single GPU
- Multilingual chat with 140+ languages
- Long-document summarisation at 128k
- Open fine-tuning and LoRA adaptation
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MMLU (5-shot) | ≈60% | 2025-03 |
| MMMU | ≈49% | 2025-03 |
| LMSYS Arena (post-release) | top small-model tier | 2025-03 |
Frequently asked questions
What is Gemma 3 4B?
Gemma 3 4B is a 4B-parameter open small LLM from Google DeepMind, released March 2025 with 128k context, image input, and 140+ language coverage.
How does it differ from Gemma 2 4B?
Gemma 3 adds multimodal vision input, extends context from 8k to 128k, improves multilingual coverage, and is distilled from newer Gemini 2.0 research.
Is Gemma 3 commercial-use friendly?
Yes — the Gemma licence permits research and commercial use, with a few restrictions similar to Llama's acceptable-use policy.
Sources
- Google DeepMind — Gemma 3 — accessed 2026-04-20
- Gemma 3 technical report — accessed 2026-04-20