What is Gemma 3 270M?
Gemma 3 270M is the smallest model in Google's Gemma 3 family: a text-only model with 268M parameters from Google DeepMind, built from the same research and technology behind the Gemini models. The unusual part is the training budget. Google trained it on 6 trillion tokens, three times the 2 trillion the 1B sibling got and more than the 4B's 4 trillion. That much training behind so few parameters is why it matters locally: it runs on effectively any machine you own.
| Specification | Gemma 3 270M |
|---|---|
| Total parameters | 268M (0.27B) |
| Architecture | Gemma 3, text-only variant |
| Context window | 32K tokens |
| Modalities | Text in, text out |
| Training data | 6 trillion tokens, knowledge cutoff August 2024 |
| Languages | Trained on content in over 140 languages |
| Release date | August 5, 2025 |
| License | Gemma |
The training mix is web documents, code and mathematics, with content in over 140 languages and a knowledge cutoff of August 2024. Unlike the 4B, 12B and 27B members of the family, the 270M takes no image input, and its context window is 32K tokens rather than the larger models' 128K. Google pitches the family at environments with limited resources, laptops, desktops or your own cloud infrastructure, and the 270M is the far end of that scale: the unquantized F16 GGUF of the instruction-tuned model is a 0.54 GB file.
Gemma 3 270M benchmarks
Google's published numbers, from the Gemma 3 model card, cover the pre-trained (PT) and instruction-tuned (IT) checkpoints of the 270M. The vendor table compares it against no other models, so read these as a capability floor for the size, not a leaderboard:
| Benchmark | Gemma 3 270M PT | Gemma 3 270M IT |
|---|---|---|
HellaSwag Commonsense completion | 40.9 | 37.7 |
BoolQ Binary questions | 61.4 | - |
PIQA Physical commonsense | 67.7 | 66.2 |
TriviaQA Trivia recall | 15.4 | - |
ARC-c Science questions | 29.0 | 28.2 |
WinoGrande Commonsense reasoning | 52.0 | 52.3 |
BIG-Bench Hard Hard reasoning | - | 26.7 |
IF Eval Instruction following | - | 51.2 |
The scores land where a 268M-parameter model should: reasonable physical commonsense (PIQA 67.7), weak hard reasoning (BIG-Bench Hard 26.7), and almost no stored world knowledge (TriviaQA 15.4). The useful number is IF Eval at 51.2: the instruction-tuned checkpoint holds up on instruction following, which matches Google's own guidance that these models do best on tasks framed with clear prompts and instructions rather than open-ended conversation.
Gemma 3 270M hardware requirements
The system requirement to check is memory, and here it barely registers. The builds come from community repos: ggml-org/gemma-3-270m-GGUF carries a single Q8_0 and unsloth/gemma-3-270m-it-GGUF carries the full ladder. The sizes below are the unsloth files of the instruction-tuned model.
| Memory | Build to pick | File size |
|---|---|---|
| 2 GB | Q4_K_M | 0.25 GB |
| 4 GB | Q8_0 | 0.29 GB |
| 8 GB and up | F16 | 0.54 GB |
The entire ladder spans 0.18 GB to 0.54 GB, so a low-bit quant saves a couple hundred megabytes at most: take Q8_0 or the F16 file and keep the quality. If the format is new to you, start with what GGUF is.
How to run Gemma 3 270M 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 270M 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 the next size up, Gemma 3 1B.
Gemma 3 270M license
Gemma 3 270M ships under Google's Gemma license, not Apache 2.0. The weights are open for both the pre-trained and instruction-tuned variants, with use governed by Google's Terms of Use and the Gemma Prohibited Use Policy. One practical note: the Hugging Face repo is gated, so you log in and accept Google's usage license, and access is granted immediately.