Vision
Google says a lightweight embedding module replaces a conventional vision encoder, letting the language-model backbone perform visual processing.
Google DeepMind / Gemma 4 12B / released June 3, 2026
Quick answer
Gemma 4 12B brings native audio to a laptop-sized tier.
Google released an encoder-free multimodal 12B model with a stated 16GB VRAM or unified-memory target, Apache 2.0 licensing, and support across local inference ecosystems.
Unified architecture
Audio
Google says raw audio is projected into the same dimensional space as text tokens without a separate audio encoder.
Text
Text, vision, and audio meet inside a unified backbone, which is the defining encoder-free architecture claim for this release.
Four checks before a local agent test
Memory
Start from Google's 16GB VRAM or unified-memory target, then budget for the runtime, quantization, KV cache, inputs, tools, and operating system.
Checkpoint
Choose the exact pre-trained or instruction-tuned artifact and review its files, notices, and model card.
Runtime
Google lists LM Studio, Ollama, AI Edge tools, Transformers, llama.cpp, MLX, SGLang, and vLLM as ecosystem paths; support and performance differ.
License
The launch says Apache 2.0. Review the exact distribution, dependencies, acceptable-use requirements, and deployment obligations.
Primary source
Introducing Gemma 4 12B Google's June 3, 2026 architecture, hardware, license, checkpoint, and ecosystem record. Gemma 4 family Compare E4B, 12B, and 26B MoE without transferring checkpoint-specific claims. Local agent guide Turn the source facts into a bounded setup and evaluation plan. Gemini 3.1 Compare a hosted Google model route with a separate product boundary. Kimi K2.5 Compare another multimodal reasoning route.
Gemma 4 12B questions, answered
Gemma 4 12B is Google's mid-sized open multimodal model released on June 3, 2026. It sits between the edge-friendly E4B and larger 26B MoE family tiers and is designed for local multimodal and agentic work.
Google's design sends vision and audio representations directly into the language-model backbone instead of relying on separate conventional multimodal encoders. Vision uses a lightweight embedding module, while raw audio is projected into the text-token space.
Yes. Google calls it the first mid-sized Gemma 4 model with native audio input and describes audio processing without a separate audio encoder.
Google targets 16GB of VRAM or unified memory, but that is not a guarantee for every laptop or workload. Quantization, context length, cache, image and audio inputs, runtime, and operating-system overhead all affect fit.
Google links pre-trained and instruction-tuned checkpoints on Hugging Face and Kaggle and names several local and production deployment ecosystems. Use Google's current model documentation for exact artifacts and setup.
This route does not claim Flowith access. Open weights and third-party runtime support do not establish a live Flowith model mapping.