As we speak, we’re introducing Gemma 4 12B, our newest mannequin designed to carry agentic multimodal intelligence on to laptops. Bridging the hole between our edge-friendly E4B and our extra superior 26B Combination of Consultants (MoE), Gemma 4 12B packages highly effective capabilities inside a lowered reminiscence footprint. It’s also our first mid-sized mannequin to characteristic native audio inputs.
Due to the developer group, Gemma 4 fashions have now crossed 150 million downloads. You’ve constructed the whole lot from wearable robotic arms for bodily help to enterprise-grade AI safety. We’re excited to see what you construct with this newest addition.
Right here’s an outline of what makes Gemma 4 12B distinctive:
- Novel unified structure: No multimodal encoders. The imaginative and prescient and audio inputs stream instantly into the LLM spine.
- Superior reasoning: Benchmark efficiency nearing our 26B mannequin, unlocking highly effective multi-step reasoning and agentic workflows.
- Laptop computer prepared: Sufficiently small to run domestically with simply 16GB of VRAM or unified reminiscence.
- Open and accessible: Launched underneath an Apache 2.0 license with help throughout the developer ecosystem.
- Drafter-ready: Gemma 4 12B comes geared up with Multi-Token Prediction (MTP) drafters to scale back latency.
Collectively, these options carry superior multimodal capabilities to on a regular basis {hardware} with out sacrificing velocity or reasoning. Let’s now take a better take a look at how Gemma 4 12B achieves this.
Run state-of-the-art brokers domestically
Gemma 4 12B delivers efficiency nearing our bigger 26B MoE mannequin on normal benchmarks, however at lower than half the whole reminiscence footprint. Sufficiently small to run domestically on shopper laptops with 16GB of RAM, it unlocks highly effective multimodal and agentic experiences proper in your machine.
