Google DeepMind has expanded its Gemma 4 lineup with a fifth model, the Gemma 4 12B, released on June 3, 2026. This is the first and only model in the family to adopt an encoder-free architecture, abandoning the dedicated vision and audio encoders that the other four sizes rely on. Designed to run entirely on local hardware with just 16GB of unified memory or VRAM, the 12B model processes text, images, audio, and video natively under the permissive Apache 2.0 license — and it is available immediately.
Why Google Dropped the Encoder
Traditional multimodal models rely on specialized encoders to translate images or audio into representations the language model can understand. The other Gemma 4 models are no exception: the 26B A4B MoE model carries a vision encoder with roughly 550 million parameters, while the edge-focused E2B and E4B include a dedicated audio encoder of about 300 million parameters. In such designs, the LLM sits idle waiting for the encoder to finish processing — a cost in both latency and memory.
The Gemma 4 12B eliminates this bottleneck entirely. For vision, it replaces the multi-layer Transformer-based vision encoder with a lightweight embedding module of just 35 million parameters. Images are divided into 48×48 pixel patches, and a single linear projection maps each patch directly into the LLM’s hidden space. Spatial information is injected by adding learned X/Y coordinate vectors — no attention layers, no cross-modal complexity. The entire operation that required a 550M-parameter encoder in other models is now handled by a minimal matrix multiplication, with each patch processed independently.
Audio processing is even more streamlined. The traditional audio encoder is gone. Instead, raw 16kHz waveforms are sliced into 40-millisecond segments (640 samples per slice) and linearly projected straight into the same token embedding space used for text. Audio enters the LLM as language — the same as text.
This design carries a major side benefit: because vision, audio, and text now share the same weight space, fine‑tuning with LoRA or full‑parameter training updates all three modalities simultaneously through a single pass. No more freezing the encoder while tuning the LLM, or managing two separate optimizers. The entire model can be adapted in one go.
Where the 12B Sits on Performance
Google positions the 12B as delivering “performance approaching a 26B MoE model at less than half the memory footprint.” Third‑party benchmarks circulating as of June 4, 2026, suggest the model scores 77.2% on MMLU Pro, 78.8% on GPQA Diamond, and 77.5% on AIME 2026 — numbers that indeed place it between the 26B A4B (82.6% / 82.3% / 88.3%) and the smaller E4B (69.4% / 58.6% / 42.5%) on these measures. The previous generation Gemma 3 27B, by comparison, scored 67.6% on MMLU Pro, 42.4% on GPQA Diamond, and just 20.8% on AIME 2026.
It is important to note that Google’s official model card has not yet published 12B‑specific benchmark scores as of June 4. The figures above come from the community and are not yet confirmed by the company. The directional picture — that the 12B competes with models twice its active parameter count — is clear, but exact numbers await official disclosure.
16GB Is the Real Target
At Q4 quantization, the model weighs approximately 6.7GB. Even with KV cache overhead, it fits comfortably within 16GB of unified memory or VRAM. That means a MacBook Pro with Apple Silicon or a notebook equipped with an NVIDIA GeForce RTX 4060 or higher can run the model locally. For the first time, Google DeepMind also ships a native macOS desktop application — Google AI Edge Gallery — that runs fully offline, handling voice and camera input without any network connection. Major inference frameworks including llama.cpp, Ollama, LM Studio, MLX, vLLM, and SGLang support the model from day one.
Google also bundles a Multi‑Token Prediction (MTP) draft model — a small auxiliary model that guesses the next several tokens ahead, allowing the main model to verify and accept them in one shot. This speculative decoding technique boosts token generation speed without degrading output quality.
First Mid‑Size Model With Native Audio
Until now, only the edge‑focused E2B and E4B models in the Gemma 4 family supported speech input natively. The 12B becomes the first mid‑size Gemma model to handle audio directly. It can perform automatic speech recognition, speaker diarization, and voice translation. It accepts audio clips up to 30 seconds and video clips up to 60 seconds at 1 frame per second. In a demo detailed on the official Google blog, the team fed a five‑minute clip from the Google I/O keynote and analyzed 313 frames and the accompanying audio simultaneously — all on a single local machine.
Gemma 4 Family Now Complete
The addition of the 12B fills a gap that had been noticeable since the April 2, 2026, launch of the Gemma 4 family. At that time, the lineup skipped the 12B size, leaving owners of the popular Gemma 3 12B without a direct upgrade path. The new model sits between the E4B (4.5B effective parameters) and the 26B MoE, plugging the hole. However, the 12B is not a simple resizing — its encoder‑free design is unique within the family, representing a structural experiment rather than a straightforward parameter bump.
The full family now spans five sizes: E2B (for smartphones), E4B (for edge devices), 12B (for notebooks and laptops), 26B A4B MoE (for consumer GPUs), and 31B Dense (for workstations and cloud instances). Total downloads across the family have surpassed 150 million since April.
The local LLM market is increasingly crowded. Qwen 3.5 and 3.6 hold ground for coding and agentic use cases, while Mistral Small 4 competes on cost efficiency. The Gemma 4 12B’s key differentiators are its permissive Apache 2.0 license, the breadth of the family spanning device classes, and now an architecture that eliminates the encoder overhead — a design choice that will be put to the test by the community in the coming weeks. Whether the trade‑off (sacrificing the richly pretrained encoder representation in exchange for simpler training and faster inference) pays off in real‑world tasks will become clear as independent evaluations and user reports emerge. For developers who want a single model that handles text, images, audio, and video inside 16GB of RAM, the Gemma 4 12B is the first option of its kind.
