Curiosity · AI Model
Gemma 3 1B
Gemma 3 1B is the smallest member of Google's Gemma 3 open-weights family, launched in 2025. At roughly a billion parameters, it is small enough to run comfortably on modern laptops and phones while still inheriting the instruction-tuning recipe and safety filters used across the Gemini-adjacent Gemma line.
Model specs
- Vendor
- Family
- Gemma 3
- Released
- 2025-03
- Context window
- 32,768 tokens
- Modalities
- text
Strengths
- Tiny footprint runs well on CPU and consumer GPUs
- Strong safety filtering inherited from the Gemma recipe
- Open weights under the Gemma license — fine-tuning friendly
Limitations
- Small capacity limits reasoning and coding ability
- Not a replacement for Gemini Pro / Flash in production
- Open-weights license has restrictions on downstream commercial use
Use cases
- On-device chat and assistant features
- Fine-tuning for classroom and research projects
- Embedded LLM use cases with tight memory budgets
- Edge-device RAG assistants
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MMLU | ≈45% | 2025-03 |
| HellaSwag | ≈70% | 2025-03 |
Frequently asked questions
What is Gemma 3 1B?
Gemma 3 1B is the smallest variant of Google's Gemma 3 open-weights family — a roughly 1-billion-parameter LLM tuned for on-device and edge deployments.
Where can I download Gemma 3 1B?
Gemma 3 1B weights are published on Kaggle and Hugging Face under the Gemma license, with Google providing reference inference code and ready-made quantised variants.
What is Gemma 3 1B good for?
On-device chat, edge RAG, classroom fine-tuning, and lightweight assistant features where you cannot afford a cloud round-trip.
Sources
- Google — Gemma 3 launch — accessed 2026-04-20
- Hugging Face — Gemma 3 1B — accessed 2026-04-20