What is Gemma 3 1B?
Gemma 3 1B is the smallest instruction-tuned model in Google's Gemma 3 family, published on Hugging Face as google/gemma-3-1b-it and built by Google DeepMind on the same research and technology used to create the Gemini models. Google's card calls these lightweight open models whose relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure. The 1B is the smallest of the four sizes Google published, at 999,885,952 parameters in the safetensors listing. The community GGUF builds in the file listing run from 0.56 GB for the smallest low-bit files to 2.01 GB for BF16.
| Specification | Gemma 3 1B |
|---|---|
| Parameters | 1B (999,885,952 weights in the safetensors listing) |
| Architecture | Gemma 3, text variant (gemma3_text) |
| Context window | 32K tokens input, 8,192 tokens output |
| Modalities | Text in, text out |
| Training data | 2 trillion tokens: web documents, code, mathematics |
| Release date | March 10, 2025 |
| License | Gemma, gated repo on Hugging Face |
This size is the text-only member of the family: the repo is tagged gemma3_text, and Google's image-understanding benchmarks start at the 4B size. You get a 32K token input window (the 4B, 12B and 27B get 128K) and up to 8,192 tokens of output per response. Training ran on 2 trillion tokens of web documents, code and mathematics, with web content in over 140 languages.
Gemma 3 1B benchmarks
Google's model card reports benchmark numbers for the pre-trained (PT) checkpoints, and open weights ship for both the pre-trained and instruction-tuned variants. Only two of its tables carry a 1B column: reasoning and factuality, and multilingual. The STEM and code and multimodal tables start at 4B. The rows below are from the reasoning and factuality table:
| Benchmark | Gemma 3 PT 1B | Gemma 3 PT 4B | Gemma 3 PT 12B | Gemma 3 PT 27B |
|---|---|---|---|---|
HellaSwag Commonsense completion | 62.3 | 77.2 | 84.2 | 85.6 |
PIQA Physical commonsense | 73.8 | 79.6 | 81.8 | 83.3 |
ARC-c Harder science | 38.4 | 56.2 | 68.9 | 70.6 |
TriviaQA Factual recall | 39.8 | 65.8 | 78.2 | 85.5 |
WinoGrande Pronoun resolution | 58.2 | 64.7 | 74.3 | 78.8 |
BIG-Bench Hard Hard reasoning | 28.4 | 50.9 | 72.6 | 77.7 |
DROP Reading comprehension | 42.4 | 60.1 | 72.2 | 77.2 |
Scaling is uneven in a useful way. On commonsense tasks the 1B keeps a workable share of the 27B's score, 73.8 against 83.3 on PIQA. Factual recall and hard reasoning shrink much faster: TriviaQA falls from 85.5 to 39.8 and BIG-Bench Hard from 77.7 to 28.4. Google's multilingual table tells the same story, the 1B scores 2.04 on MGSM against 74.3 for the 27B, so keep it on English-leaning work. The card itself notes these models are not knowledge bases: use the 1B for drafting, summarization and chat, and check anything factual it tells you.
Gemma 3 1B hardware requirements
The system requirement to check is memory, and at this size almost any machine passes. The builds below are from the community repo unsloth/gemma-3-1b-it-GGUF, with the real file sizes from its listing. That repo also carries low-bit files if you want to go smaller: UD-IQ1_M and UD-IQ1_S are 0.56 GB each, Q2_K is 0.69 GB, and the intermediate Q5_K_M lands at 0.85 GB.
| Memory | Build to pick | File size |
|---|---|---|
| 4 GB | Q4_K_M | 0.81 GB |
| 8 GB | Q8_0 | 1.07 GB |
| 16 GB and up | BF16 | 2.01 GB |
When two builds both fit, take the larger one. Google also ships its own quantized build in the repo google/gemma-3-1b-it-qat-q4_0-gguf, a single Q4_0 file at 1.00 GB. The largest file in the unsloth listing is BF16 at 2.01 GB; the largest quantized builds there are UD-Q8_K_XL at 1.48 GB and Q8_0 at 1.07 GB. If the format is new to you, start with what GGUF is.
How to run Gemma 3 1B in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for Gemma 3 1B in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
For the rest of the family, see every Gemma model you can run locally, or go even smaller with Gemma 3 270M.
Gemma 3 1B license
Gemma 3 1B is released under Google's Gemma license. The Hugging Face repo is gated: you log in, review the usage terms and accept them, and access is granted immediately. What you can do with the weights is set by the Gemma Terms of Use, and prohibited uses are listed in the Gemma Prohibited Use Policy, so read both before building the model into a product.