gemma-3-270m

Updated
03.07.2026
Reasoning
Multilingual
Run

Run Gemma 3 270M, an ultra-light Google model, locally in Atomic Chat. Fast on-device inference, fully offline and private. Free.

huggingface-cli download google/gemma-3-270m
from transformers import AutoModel
model = AutoModel.from_pretrained("google/gemma-3-270m")
More models
View all

  
  

NameSize / UsageContextInput
gemma-4-31B-it
256KText, Image, Audio
gemma-4-26B-A4B-it
256KText, Image, Audio
gemma-4-12B-it
125KText, Image, Audio
diffusiongemma-26B-A4B-it
256KText, Image, Audio
gemma-4-E2B-it
125KText, Image, Audio
gemma-3-1b-it
32KText
At a glance
  • License: Gemma
  • Context length: 32K tokens
  • Languages: Multilingual
  • Minimum hardware: ~2 GB VRAM
  • Strengths: ultra-light tasks and fine-tuning base
Overview
gemma-3-270m is the smallest member of Google's Gemma 3 family, a text-only model with just 0.27B parameters.
  • License: Gemma
  • Context length: 32K tokens
  • Languages: Multilingual
  • Minimum hardware: ~2 GB VRAM
  • Strengths: ultra-light tasks and fine-tuning base
Highlights
  • 0.27B-parameter model sized for local hardware.
  • Runs on ~2 GB VRAM.
  • Handles long inputs with a 32K-token context window.
  • Strong at structured extraction tasks.
  • Also handles multilingual text handling out of the box.
What it is good at
gemma-3-270m carries reasoning and multilingual capabilities, and at this size it shines on focused, repeatable jobs rather than open-ended conversation:
  • Structured extraction — pulling entities, fields, or sentiment out of raw text and returning clean structured output, the kind of task Google highlights for this model.
  • Multilingual text handling — the 256k vocabulary covers rare tokens and many languages, so it can route, classify, or rewrite text across languages on-device.
  • Fine-tuned task agents — as a fine-tune base it is cheap to specialize for one job (query routing, tagging, compliance checks) and then run that specialized version locally.
Running it locally
With 0.27B parameters and a 32K context window, gemma-3-270m is one of the lightest LLMs you can run. Quantized to INT4 it needs under 200 MB of memory, which means it fits on laptops, desktops, and even phones without a dedicated GPU. Pull the weights from Hugging Face:
huggingface-cli download google/gemma-3-270m
You can load it through Transformers or vLLM for scripting and serving, or open it in Atomic Chat with one click and start chatting on-device, no setup beyond the download.
License
gemma-3-270m is released under the Gemma license. It provides open weights and permits responsible commercial use, including fine-tuning and deploying your adapted version in your own projects, subject to Google's usage terms.
Desktop



macOS
(M1 or better)
Download
Windows
(x64)
Download
Linux
(x86_64)
Download
Frequently asked questions
gemma-3-270m is the smallest model in Google's Gemma 3 family, a text-only transformer with 0.27B parameters and a 32K context window. It is a fine-tune base built for narrow, efficient tasks like extraction, classification, and routing rather than broad chat. You can run it fully on-device through Atomic Chat.
Very little. Quantized to INT4, gemma-3-270m uses under 200 MB of memory and runs without a dedicated GPU, so a standard laptop, desktop, or even a phone can handle it. For local fine-tuning rather than inference, you'd want an NVIDIA GPU with around 8 GB of VRAM, but plain inference in Atomic Chat works on modest CPUs.
Yes. The weights are open and free to download from Hugging Face under the Gemma license, which permits responsible commercial use including fine-tuning and deploying your own version. Running it locally in Atomic Chat costs nothing beyond your own hardware and electricity.
Yes. Once you download the weights, gemma-3-270m runs entirely on your machine with no network connection. In Atomic Chat your prompts and outputs stay on-device, which keeps the model usable on a plane, in the field, or anywhere private data shouldn't leave your computer.
It is strongest on focused, high-volume tasks: entity extraction, sentiment analysis, query routing, and turning unstructured text into structured output. As a fine-tune base it is cheap to specialize for one job and then run that version locally. For long open-ended conversation or deep reasoning, a larger model is a better fit.