# Gemma 3 1B ![Organization](https://img.shields.io/badge/Org-Google%20DeepMind-4285F4) ![Type](https://img.shields.io/badge/Type-Open%20weights-555) ![Context](https://img.shields.io/badge/Context-32K%20tokens-1f6feb) ![Parameters](https://img.shields.io/badge/Params-1B-8957e5) ![License](https://img.shields.io/badge/License-Gemma%20(open)-2da44e) ![Modality](https://img.shields.io/badge/Modality-Text-555) ![Released](https://img.shields.io/badge/Released-Mar%202025-555) > **Quick answer:** Gemma 3 1B is the smallest model in Google DeepMind's Gemma 3 family, released on 12 March 2025 with fully open weights under the permissive Gemma license. It has 1 billion parameters, a 32,768-token context window, and handles text-only input and output. Trained on 2 trillion tokens across 140+ languages, it is designed to run on CPU without a GPU and is well-suited for on-device applications, edge deployments, and lightweight fine-tuning. On MMLU-Pro it scores 14.7%, reflecting its intentionally small footprint. ## At a Glance **Where Gemma 3 1B leads** - Smallest model in a strong open family: fits on CPU, runs on mobile, laptops, and resource-constrained edge devices. - Open weights with a permissive Gemma license — commercial use allowed, no per-query cost when self-hosted. - Trained on 2 trillion tokens including code and multilingual data — capable for its size on text generation and simple QA. - Active fine-tuning ecosystem: 550+ fine-tuned community variants on HuggingFace, 1M+ monthly downloads. **Where it lags** - Text-only — no image understanding (the 4B, 12B, 27B variants are multimodal). - 32K context window vs 128K for all larger Gemma 3 sizes. - Benchmark scores are substantially below larger models: 14.7% on MMLU-Pro vs 85.2% for Gemma 3 27B. **Best for:** on-device inference, edge AI, low-resource environments, fine-tuning experiments, and applications where model size and cost matter more than frontier accuracy. ## What Gemma 3 1B Is Gemma 3 1B is part of Google's Gemma 3 family of open-weight models, which was built from the same research and technology stack that underlies the proprietary Gemini 2.0 models. The 1B variant is the entry-level member of the family, designed specifically for deployment in environments where compute is limited — phones, laptops, IoT devices, and offline applications. Unlike the 4B, 12B, and 27B Gemma 3 models — which are multimodal and can process both images and text — Gemma 3 1B is text-only. Its context window is also shorter at 32,768 tokens, compared to 128K for the rest of the family. These trade-offs are deliberate: the model is optimized for minimal footprint and fast inference on CPU rather than for top-end accuracy. The model ships in two variants: a base pre-trained model (`gemma-3-1b-pt`) and an instruction-tuned model (`gemma-3-1b-it`). The instruction-tuned variant supports function calling and structured output, enabling simple agentic workflows even on constrained hardware. It is compatible with Hugging Face Transformers (4.50+), Ollama, vLLM, and Google AI Edge for on-device deployment on Android and web. ## Specifications | Field | Value | |---|---| | Organization | Google DeepMind | | Parameters | 1 billion | | Architecture | Transformer decoder (Gemma 3) | | Modality | Text only (no vision) | | Context window | 32,768 tokens | | Max output | 8,192 tokens | | Training tokens | 2 trillion | | Training languages | 140+ | | License | Gemma (permissive, commercial use allowed) | | Release date | 12 March 2025 | | Open weights | Yes — HuggingFace, Kaggle, Ollama | | HuggingFace ID | [`google/gemma-3-1b-it`](https://huggingface.co/google/gemma-3-1b-it) | | Inference frameworks | Transformers, vLLM, Ollama, JAX, Keras, Google AI Edge | ## Pricing Gemma 3 1B has open weights — there is no per-token charge when self-hosted. Cloud API pricing varies by provider: | Provider | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | Self-hosted (open weights) | Free | Free | | Google AI Studio | Free (rate-limited) | Free | | Vertex AI | Usage-based | Usage-based | ## Benchmark Results | Benchmark | Score | Notes | |---|---|---| | MMLU-Pro | 14.7% | Very low — expected for 1B scale | | HellaSwag (10-shot) | 62.3% | Commonsense reasoning | | ARC-Challenge (25-shot) | 38.4% | Grade-school science | | BIG-Bench Hard | 28.4% | Multi-step reasoning | | MATH (4-shot) | 24.2% | Mathematical reasoning | | GSM8K (8-shot) | 38.4% | Grade-school math | | MBPP (3-shot) | 46.0% | Python code generation | | HumanEval (0-shot) | 36.0% | Function synthesis | | MGSM (multilingual math) | 2.04 | Limited multilingual math | *Scores from the Gemma 3 technical report (March 2025). MMLU-Pro from [Benchgen leaderboard](/benchmarks/mmlu-pro).* ## Gemma 3 1B vs Other Small Open Models | Model | Params | Context | MMLU-Pro | Modality | License | |---|---|---|---|---|---| | **Gemma 3 1B** | 1B | 32K | 14.7% | Text | Gemma (open) | | [Gemma 3 4B](/models/google/gemma-3-4b) | 4B | 128K | 43.6% | Text + Image | Gemma (open) | | Qwen3.5-0.8B | 0.8B | — | 42.3% | Text | Apache 2.0 | | Qwen3.5-2B | 2B | — | 66.5% | Text | Apache 2.0 | | Phi 4 Mini | 4B | — | 52.8% | Text | MIT | | Llama 3.1 8B | 8B | 128K | 48.3% | Text | Llama 3.1 | Gemma 3 1B scores below models of similar or larger size on MMLU-Pro, which reflects the extreme parameter budget. Within the same family, Gemma 3 4B (4× the parameters) scores nearly 3× higher at 43.6%. For tasks where raw benchmark performance matters, stepping up to 4B or 12B is recommended. ## Run Gemma 3 1B on Benchgen Gemma 3 1B's published benchmark numbers reflect Google's evaluation setup. Running it on Benchgen lets you measure performance on your specific task domain, prompt format, and input distribution — giving you comparable, version-tracked scores for your actual use case rather than general academic benchmarks. Useful when evaluating whether the 1B model is sufficient for a given application, or to justify upgrading to a larger variant. ## Related Models & Benchmarks - Same family: [Gemma 3 4B](/models/google/gemma-3-4b) · [Gemma 3 12B](/models/google/gemma-3-12b) · [Gemma 3 27B](/models/google/gemma-3-27b) - Comparable size: Qwen3.5-0.8B · Phi 4 Mini - Benchmarks: [MMLU-Pro](/benchmarks/mmlu-pro) · [SWE-bench Verified](/benchmarks/swe-bench-verified) ## Frequently Asked Questions

What is Gemma 3 1B?

Gemma 3 1B is the smallest model in Google DeepMind's Gemma 3 open-weight family, released on 12 March 2025. It has 1 billion parameters, a 32K context window, and handles text-only input and output. It is designed for on-device and resource-constrained deployments where a larger model would not fit.

Does Gemma 3 1B support image input?

No. Gemma 3 1B is text-only. The multimodal (image + text) capability is available in the 4B, 12B, and 27B sizes of the Gemma 3 family, but was not included in the 1B variant due to its size constraints.

Can Gemma 3 1B run on CPU?

Yes — this is one of its main design goals. At 1B parameters with BF16 weights (~2 GB), it fits in RAM on most laptops and desktops and can run on CPU via Gemma.cpp, Ollama, or llama.cpp without requiring a GPU.

What is the Gemma license?

The Gemma license is a custom permissive license from Google that allows commercial use and redistribution. It prohibits using the model to build products that compete directly with Google's Gemini API services, and requires attribution. It is not Apache 2.0, but is broadly permissive for most commercial applications.

How does Gemma 3 1B compare to Gemma 3 4B?

Gemma 3 4B has 4× the parameters, a 128K context window (vs 32K), and adds multimodal image understanding. On MMLU-Pro it scores 43.6% vs 14.7% for 1B. The 4B model is substantially more capable for reasoning and knowledge tasks; the 1B is better suited for cases where model size or inference cost is the primary constraint.

--- *Benchmark scores from the [Gemma 3 technical report](https://goo.gle/Gemma3Report) (Google DeepMind, March 2025) and Benchgen leaderboard data. Last updated 2026-06-18.*

Gemma 3 1B — Evaluations

Evaluations