Gemma 4 12B
2026-06-03 · Publisher : Google · Domain : Artificial Intelligence
Google DeepMind adds Gemma 4 12B to the Gemma 4 family: a mid-size open model, without separate encoders for vision and audio, that runs on a 16 GB laptop.
What changes
- Unified encoder-free architecture: visual and audio inputs go directly into the language model.
- First mid-size open model in the family with native audio input.
- Performance close to the 26B MoE for less than half the memory footprint.
- Multi-token prediction (MTP) drafters provided to reduce latency.
- Gemma 4 models exceed 150 million downloads.
### Specifications - Size: about 12 billion parameters (11.95B), 256K tokens of context. - Modalities: text, image and audio as input. - Availability: open weights (Hugging Face, Kaggle), LM Studio, Ollama, llama.cpp, MLX, vLLM, SGLang. - License: Apache 2.0. ### Pricing Open weights under Apache 2.0. Free use through the Gemini API according to the pricing page. ### Benchmarks | Benchmark (model card, scores published by the vendor) | Gemma 4 12B | |---|---| | MMLU Pro | 77.2% | | AIME 2026 | 77.5% | | LiveCodeBench v6 | 72.0% | | GPQA Diamond | 78.8% | | MMMU Pro (vision) | 69.1% | ### Key takeaways A compromise between the E4B and the 26B: sufficient for local multimodal agents on a standard workstation.
Links
Tags : LLM, Open Source, Google, Gemma, Multimodal