Gemma 4 12B

Spec

Attribute	Value
Developer	Google DeepMind
Released	2026-06-03
Announced	2026-06-03
Context window	unknown
Pricing	unknown
License	Apache 2.0 (free, commercial use)
Availability	Open weights on Hugging Face and Kaggle
Parameters	12B (dense decoder-only transformer)
Architecture	Encoder-free — vision and audio inputs flow directly into LLM backbone (no separate vision/audio encoders)
Modalities	Text, image, audio, video (native unified, no encoder bridges)
Deployment	vLLM, SGLang, MLX, llama.cpp compatible

Release Date

2026-06-03

Local Requirements

• Full precision: 16GB VRAM (RTX 4060, RTX 5090, Apple Silicon Mac)
• Quantized: 8GB VRAM
• No cloud compute required for inference

Benchmarks

All figures from Google internal evaluation; independent third-party benchmarks pending as of June 4, 2026.

Benchmark	Score	Notes
GPQA Diamond	78.8	Graduate-level reasoning
MMLU Pro	77.2%	Broad knowledge
Vision	strong	No separate encoders
Coding	competitive	Agentic function calling

• Approaches the larger Gemma 26B MoE model at roughly half the memory footprint
• Supports optional thinking mode (step-by-step CoT reasoning)

Features

• Function calling: text and multimodal function calling for local agentic workflows
• Thinking mode: chain-of-thought style reasoning on demand
• Agentic use: plans, navigates apps, completes multi-step tasks locally

Use Cases

• Local AI agent workflows (no cloud dependency)
• Multimodal analysis: image captioning, audio transcription, video understanding on-device
• Privacy-sensitive enterprise deployments
• Developer experimentation: fine-tuning on 8GB+ VRAM

Compared To

Model	Params	License	Multimodal	Local 16GB	Notes
Gemma 4 12B	12B	Apache 2.0	text+image+audio+video	✅	unified encoder-free
Gemma 3n (prev)	5-8B	Apache 2.0	text+image	✅	mobile-first (PLE arch)
Llama 4 Scout	~17B active	Llama license	text+image	⚠️ partial	Meta MoE
Mistral Small	22B	Apache 2.0	text	✅ 8GB+	text only
Phi-4	14B	MIT	text+image	✅	Microsoft

Significance

• Encoder-free multimodal: a novel unified architecture eliminates audio/vision encoder overhead — enabling true laptop-class all-modal AI
• Open-source frontier push: Apache 2.0 multimodal at 12B is the most capable open-weights on-device model at this capability level as of June 2026
• Local agents: directly enables local autonomous agent workflows that previously required cloud inference
• Follows Gemma 3n (May 2026, mobile-first) — Gemma 4 12B is the full laptop/desktop tier of Google's open model strategy

Sources

• Introducing Gemma 4 12B — Google Blog (snapshot)
• Gemma 4 model page — DeepMind
• Google AI Edge dev guide

Open Questions

• Independent third-party benchmarks on consumer hardware?
• Are Gemma 4 27B / larger variants planned?
• Fine-tuning guides and instruction-tuned variants?