Breaking

Google Releases Gemma 4 12B, a Unified Multimodal Model for Laptops

June 14, 2026 at 02:11 EDT

Foundation Models
Open Source
Media Generation

Google DeepMind on June 3, 2026 released "Gemma 4 12B," a unified multimodal model that natively processes text, images, audio and video, under a commercially usable Apache 2.0 license. Designed to run on 16GB-class laptops while delivering high-performance reasoning, it is immediately available via Hugging Face, Ollama, Google AI Edge, LM Studio and Kaggle. The official announcement positions it as a model that "bridges the gap" between edge efficiency and advanced reasoning.

June 3, 2026 · Google DeepMind

Gemma 4 12B: a multimodal open model that fits on your laptop

A unified, encoder-free model handling text, images and audio/video natively — running locally on a typical 16GB machine, released as open weights under Apache 2.0.

11.95B

Parameters (Dense)

256K

Max context tokens

16GB

RAM/VRAM with quantization

Modalities: text · image · audio

Encoder-free: same performance, half the memory

Dropping separate vision/audio encoders, it approaches the larger 26B MoE model on under half the memory footprint.

26B MoE

full footprint

Gemma 4 12B

< half the memory

Benchmark scores

GPQA Diamond78.8

LiveCode Bench72

BBEH — hard reasoning53

What developers like

Strong local agent on Mac: autonomous code, tool use, dependency resolution
256K context practical in a 16GB setup
Encoder-free design makes video/audio smooth; QAT variant praised

Caveats noted

~44 tokens/s on Mac M5 Pro (24GB) seen as slow
Trails the Qwen 3.6 family in some cases; narrower knowledge breadth
CPU-only setups need tuning with llama-server and tool orchestration

Filling the middle

The 12B slots between the edge-focused E2B/E4B and the larger 26B MoE / 31B Dense models — pairing mobile-class efficiency with advanced reasoning, and competing directly with mid-size Llama 3.2, Mistral and Qwen-family local models.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.