Gemma 4 12B: Google's Unified Open Multimodal Gemma 4 Model
Type: AI model · Identifier: 4 (12B Unified, 11.95B parameters) · Operator: Google DeepMind (claimed, unverified)
Gemma 4 12B Unified is a Google DeepMind open-weights model in the Gemma 4 family: 11.95 billion parameters, a 256K-token context, and native text, image and audio input.
What is documented, and what is not
| Operator | Google DeepMind (claimed, unverified) · Official documentation |
|---|---|
| User-Agent | Not a web visitor: a model has no User-Agent of its own. Software built on it may. |
| IP ranges | No dedicated official IP range: a model is not a network crawler. Network behavior of apps that use it is unverified here. |
| robots.txt | Not applicable: a model has no robots.txt token of its own. |
Data: Epoch AI, CC BY. Training-compute figures cited from that dataset are estimates.
What the official sources say about Gemma 4 12B
Gemma 4 12B Unified is one of five sizes in Gemma 4, a family of open models from Google DeepMind. The official Gemma 4 model card, last updated 30 July 2026 UTC, lists the sizes as E2B, E4B, 12B, 26B A4B and 31B, in pretrained and instruction-tuned weights. The 12B model is the "Unified" variant: Google says it has 11.95 billion parameters, 48 layers, a 1,024-token sliding window, a 262,000-token vocabulary and a 256,000-token context window. Epoch AI records "Gemma 4 12B" with a publication date of 3 June 2026, organization Google DeepMind, confidence "Confident", and accessibility "Open weights (unrestricted)." Data: Epoch AI, CC BY.
What "Unified" means is specific to this size. Google says other Gemma 4 models use dedicated encoders for multimodal data, but Gemma 4 12B drops those encoders and projects raw image patches and audio waveforms into the model's embedding space through lightweight linear layers, so all modalities go into a single decoder-only transformer. The card says this reduces multimodal latency and lets the whole model be fine-tuned in one pass. It natively accepts text, image and audio — audio is also native on E2B and E4B, and is not listed for the 31B dense model — and it generates text. Google documents automatic speech recognition and speech-to-translated-text on the audio-capable sizes, with a maximum audio length of 30 seconds and video as frames up to 60 seconds at one frame per second. Images can use variable aspect ratio and resolution.
The wider Gemma 4 design, which this size shares, uses a hybrid attention pattern that interleaves local sliding-window attention with full global attention, with the final layer always global, and Proportional RoPE on global layers. Google describes configurable thinking modes, native function calling, native system-prompt support, and multilingual use: 35 or more languages out of the box, pretrained on 140 or more. The pre-training data is described as web documents, code, images and audio, with a cutoff of January 2025, and with filtering for child sexual abuse material and automated removal of certain personal and sensitive data. The license on the card is Apache 2.0, which is a different term from earlier Gemma releases that used Google's Gemma terms; anyone citing this model should read the card's current license line rather than assume the Gemma 3 terms still apply.
This is a model, not a website visitor. Gemma 4 12B has no User-Agent, IP range or robots.txt behavior of its own, and it does not appear under its own name in server logs. What a site may see is an app or agent built on it, which should be judged by that product's documented identity. Google positions the 12B size for consumer GPUs and workstations, and notes enhanced coding and agentic capabilities. Open weights under Apache 2.0, plus on-device and local deployment of nearby sizes, mean capable text, image and audio models can run outside a provider's monitoring.
Google's card lists misuse, harmful content, privacy and bias among the risks, and points developers to the Responsible Generative AI Toolkit. For readers of FraudMeNot, the practical point is the same as for other capable open models: a fluent, personal message, including one that refers to an image or a short voice clip, is not evidence that the sender is who they claim. Confirm requests for money, codes or account changes through a contact you already know.
What to do with it
For everyday users: Gemma 4 12B is a model that people and apps run; on its own it does not send messages. Be cautious with unsolicited messages that invoke Google or Gemma and ask for payment or a login, and remember that fluent text, or a reference to an image or a short audio clip, is not proof a message is genuine.
For website owners: There is nothing to block for the model itself. Judge any bot or app built on Gemma 4 by its own documented identity. Plan for well-written, multilingual phishing, including messages that appear to react to images or short audio, that impersonate your brand.
ALLOW or DISALLOW per your AI policy; always require human step-up on login and payment routes.
Common Questions About Gemma 4 12B
Is Gemma 4 12B a crawler?
No. It is an AI model. It has no User-Agent or IP range; applications built on it may visit websites under their own names.
What is special about the 12B Unified size?
Google's model card says this size has 11.95 billion parameters and a 256K context, takes text, image and audio, and uses an encoder-free architecture that projects image and audio directly into the decoder, unlike other Gemma 4 sizes that use dedicated encoders.
What is the license?
The Gemma 4 model card states Apache 2.0. That differs from earlier Gemma terms; check the current card rather than assuming a previous license.
Sources
Sources (checked Nov. 16, 2026):, Google AI for Developers: Gemma 4 model card, Epoch AI: AI models dataset (record "Gemma 4 12B", CC BY). FraudMeNot is not affiliated with Google DeepMind (claimed, unverified). General information, not security advice.