Introducing Gemma 4 12B: a unified, encoder-free multimodal model

Google DeepMind has a new mid-size entry in its Gemma 4 family, and the pitch is refreshingly concrete: run genuinely capable multimodal AI on a laptop with 16GB of memory. Gemma 4 12B slots between the tiny edge-focused E4B and the beefier 26B Mixture of Experts model, and it's the first model of this size from Google to accept native audio input alongside vision and text.

What's actually interesting here isn't the size, it's the plumbing. Most multimodal models bolt on separate encoders that translate images or sound into something the language model can digest, and that translation step costs both memory and latency. Gemma 4 12B skips it. On the vision side, DeepMind swapped the encoder for a stripped-down embedding module built from little more than a matrix multiplication, positional embeddings, and normalization, letting the core language model handle the heavy lifting itself. Audio gets an even more aggressive trim: there's no dedicated audio encoder at all. The raw audio signal gets projected straight into the same vector space as text tokens, which is a fairly unusual design choice for a model this size.

The payoff, according to DeepMind, is performance that approaches the 26B MoE model on standard benchmarks while using less than half the memory. That's the kind of tradeoff that actually matters for people trying to build agentic tools that run locally rather than pinging a cloud API for every step. Gemma 4 12B also ships with Multi-Token Prediction drafters baked in, a technique meant to cut down response latency, which suggests DeepMind is specifically chasing snappier on-device agent workflows rather than just chasing benchmark scores.

DeepMind is releasing the whole thing under an Apache 2.0 license, with checkpoints on Hugging Face and Kaggle and support across the usual local-inference toolchain: llama.cpp, MLX, vLLM, Ollama, LM Studio, and others. There's also a new Gemma Skills Repository aimed at developers building agents on top of these models. The company says the broader Gemma 4 line has now passed 150 million downloads, and points to community projects ranging from robotic arms to enterprise security tools as evidence the open-weights strategy is paying off in actual deployment, not just leaderboard chatter.