← All models

Gemma 4 12B

Published Publisher Google

Overview

Gemma 4 12B Unified is an open-weight, dense multimodal model from Google DeepMind, designed for local use on laptops and other consumer hardware. It was released on June 3, 2026 under the Apache 2.0 license. Pre-trained and instruction-tuned checkpoints are available. (blog.google)

Architecture and capabilities

  • Size and context: 11.95 billion parameters and a 256K-token context window. (ai.google.dev)
  • Unified architecture: Instead of separate vision and audio encoders, lightweight projection layers feed image patches and audio signals into a shared decoder-only language model. (ai.google.dev)
  • Modalities: Accepts text, images, audio and video inputs and generates text. Supported tasks include speech recognition, speaker diarization, image and document understanding, video analysis, reasoning and coding. (ai.google.dev)
  • Agentic development: Supports function calling and configurable reasoning modes. A separate multi-token prediction draft model can be used for faster inference. (ai.google.dev)

Uses and deployment

Potential uses include local assistants, transcription, document analysis, coding agents and multimodal research. Google says the model can run locally on laptops with 16 GB of GPU VRAM or unified memory. Actual requirements depend on precision, inference software and workload; Google estimates approximately 26.7 GB to load the model at BF16, 13.4 GB at SFP8 and 6.7 GB at Q4_0, including its stated loading overhead. (ai.google.dev)

Limitations

The model generates text rather than audio or images. Its responses can be incorrect, outdated or biased, and complex tasks or nuanced language may be difficult. Applications should independently evaluate outputs and implement safeguards appropriate to their use. Longer contexts and different quantization settings can also affect resource needs and performance. (ai.google.dev)

Pricing

Google-hosted input and output prices for this specific 12B model: not verified, including currency and billing unit. Google's Gemini API pricing page lists a free tier for the Gemma 4 family, but its supported-model documentation lists the 31B and 26B A4B models, not the 12B model; the family-level listing therefore should not be treated as a confirmed 12B API offer. Google Cloud's pricing page lists Gemma 4 26B, not a 12B token price. Self-hosting incurs hardware and operating costs. Checked November 16, 2026. (aistudio.google.com)

Official sources