Google Drops Six AI Releases Including Gemma 4 12B Running on a Laptop

Google AI dropped one of its densest shipping weeks of the year, with six releases spanning enterprise image generation, multi-agent science, on-device multimodal models, and a live music instrument you can actually play. The throughline is a bet on agentic workflows and on-device intelligence, with most launches either generally available or open-weights from day one.

Nano Banana graduates from preview

Nano Banana 2 and Nano Banana Pro are now generally available through the Gemini Enterprise Agent Platform, Gemini API, and AI Studio. On May 28, Google DeepMind officially promoted both the gemini-3.1-flash-image-preview and gemini-3-pro-image-preview image models to General Availability, meaning developers no longer need to use temporary IDs with the -preview suffix in production environments. The preview versions will retain roughly a four-week buffer period after GA, with official retirement on June 25, so any production code still using the preview suffix must be updated before then to avoid 404 errors.

The dual-model strategy is now explicit: use Flash (Nano Banana 2) by default, and upgrade to Pro for scenarios involving text or branding. Pricing reflects that split, with Nano Banana Pro at $0.134 per 1K or 2K image and $0.24 for 4K, while Nano Banana 2 generates in under 2 seconds at roughly $0.02 to $0.04 per image. Batch and Flex pricing on Vertex cuts standard rates in half at the cost of async delivery, which adds up quickly for non-realtime workflows like nightly product image refreshes.

One quietly significant addition: Nano Banana 2 now supports video files as an input prompt, expanding multimodal conditioning beyond reference images. And the input context window for gemini-3.1-flash-image can reach 1 million tokens, which is rare for image generation models and lets you feed in large numbers of reference images and brand assets in a single call.

An idea tournament for science

Co-Scientist is Google's new multi-agent system aimed at the messiest part of research: generating hypotheses worth testing. Published in Nature, the multi-agent AI system is built with Gemini and iteratively generates, debates, and evolves novel hypotheses for complex scientific problems.

The architecture splits the work across specialized agents. A generation agent proposes initial focus areas and hypotheses grounded in literature, a proximity agent clusters them to encourage broad exploration, a reflection agent evaluates ideas for correctness and novelty, and a ranking agent compares competing ideas through an idea tournament that borrows from AlphaGo and AlphaStar but applies them to scientific debate, ranking hypotheses on potential value, robustness, and testability.

Early validation is concrete: Stanford geneticist Gary Peltz is using Co-Scientist to accelerate the search for liver fibrosis treatments, and by analyzing biomedical literature the system highlighted overlooked drug-repurposing candidates, including one that blocked 91% of a scarring-linked response in lab tests. Researchers can sign up via Hypothesis Generation, an experimental tool at labs.google/science. The honest caveat: Co-Scientist did not achieve autonomous scientific discovery, did not complete clinical trials, and did not replace a biomedical research team.

Dreambeans turns your inbox into a story feed

From Google Labs, Dreambeans is an experimental app that works overnight to curate a small daily collection of stories drawn from your connected Google apps. With your permission, Dreambeans uses Personal Intelligence to connect information from Gmail, Calendar, Photos, YouTube and Search history to curate stories. Each story is visually crafted using Nano Banana 2, and if a story involves you or people close to you, it can use Google Photos and Nano Banana 2 to paint your likeness directly into the scene instead of using generic stock imagery.

The pitch is anti-doomscroll: a finite morning drop rather than an infinite feed. Dreambeans is rolling out to eligible Google AI Ultra subscribers aged 18 and older in the U.S. on Android and iOS, with a waitlist for others.

Gemma 4 12B ditches the encoders

The biggest architectural story of the week is Gemma 4 12B, a dense 12-billion-parameter open-weights model that handles text, images, audio, and video natively without separate encoders. Traditional multimodal models rely on separate encoders to translate images and audio before passing those representations to the language model, which adds latency and increases memory usage, so Gemma 4 12B was trained with an encoder-free architecture to integrate audio and vision input directly.

In plain terms: instead of running images through a vision tower and audio through an ASR-style stack before the LLM ever sees them, raw audio waveforms and image patches are projected straight into the same token space as text. The encoder-free design reduces multimodal latency because the LLM backbone starts processing immediately, without waiting for an encoder to finish first.

Performance and footprint are the headline numbers:

• Performance nearing the larger 26B MoE Gemma model on standard benchmarks, but at less than half the total memory footprint, small enough to run locally on consumer laptops with 16GB of RAM
• Runs locally on a 16 GB laptop under an Apache 2.0 license
• First mid-sized Gemma model to feature native audio inputs
• Gemma 4 models have now crossed 150 million downloads

For developers, that unlocks a meaningful new category. Use cases include automatic speech recognition that transcribes audio natively without an external ASR pipeline, agentic reasoning that runs multi-step workflows locally with performance approaching the 26B MoE model, speaker diarization, and video understanding that processes frames alongside audio. A quick local install via Hugging Face:

pip install transformers accelerate
# Download weights from Hugging Face: google/gemma-4-12b
# Then load with native multimodal inputs (text + image + audio)

Alongside the model, Gemma 4 12B is released under Apache 2.0 and includes Multi-Token Prediction (MTP) drafters to improve inference speed and reduce latency. The drafters and base models are also optimized with Quantization-Aware Training, where the model learns during training to stay accurate when its weights are squeezed down to low-precision integers, cutting on-device memory without the usual quality hit.

Magenta RealTime 2 is a model you play

Magenta RealTime 2 is the most unusual release of the bunch: an open-weights 2.4B parameter model capable of high-quality real-time music synthesis with low-latency real-time controls via MIDI, text, and audio. MRT2 achieves roughly 15x lower latency than version one, works on standard hardware and integrates directly into DAWs.

The architectural trick is frame-level autoregression. Both MRT and MRT2 are codec language models operating on sequences of audio tokens from the SpectroStream codec, but MRT2 achieves lower latency by performing frame-level autoregression with frame-aligned conditioning, with MIDI and style prompts injected as frame-aligned conditioning at every generation step, allowing the model to react within a single 40 ms frame. End-to-end control latency lands around 200 ms, which is the difference between sending commands and improvising with the model in real time.

The Base model requires an M3 Pro or M2 Max or higher, while the Small model runs on any Apple Silicon MacBook including MacBook Air, with weights weighing about 2.5GB for base and 450MB for small. Try it from Python:

uv pip install "magenta-rt[mlx]"
mrt models init
mrt models download
mrt mlx generate --prompt "disco funk" --duration 4.0 --model=mrt2_base

Weights are released as google/magenta-realtime-2 on Hugging Face under CC BY 4.0 with code under Apache 2.0. It ships with an AUv3 plugin for DAWs and extensions for Max/MSP, PureData, and SuperCollider, which makes it less a research demo and more a working instrument for live performance, accessibility-focused music making, and interactive soundscapes in games or installations.

Taken together, the week reads as a coordinated push to put Google's models closer to the workflows they serve, whether that means SLA-backed image generation inside enterprise pipelines, a peer-reviewed hypothesis engine for biology labs, or open-weights multimodal and music models running entirely on a laptop.