Back to The Weekly Recall

More of the AI on your laptop can stay on your laptop

Google’s local image-and-audio model and Microsoft’s browser AI put more processing on the device. Plus agent controls and robotics simulation.

A seated CRT agent works on a laptop with an unplugged network cable beside it.

On June 3, Google released Gemma 4 12B, an Apache-2.0 model designed to handle images and audio locally on laptops with 16GB of memory. The model skips separate encoders, the networks that conventionally interpret images and audio before passing their outputs to a language model. Lightweight mappings feed those inputs into the language model directly.

Microsoft is also moving language work into the browser. Edge added language detection and translation and previewed speech recognition that can process audio without a cloud connection.

Google reports benchmark performance near its larger 26B mixture-of-experts model at less than half the total memory footprint. Actual runtime memory depends on the deployment configuration.

Pretrained and instruction-tuned weights are available through Hugging Face and Kaggle. Google also supplies small drafting models that propose tokens for the main model to check, reducing generation latency.

Top Stories#

Microsoft introduced Windows controls for agent access to files and networks. Microsoft Execution Containers let developers specify resource boundaries that Windows enforces during execution. The SDK is in early preview.

Windows 365 for Agents, for workflows in which agents use computers, became generally available. Microsoft also expanded Windows AI APIs beyond neural processing units to supported CPUs and GPUs.

NVIDIA announced a model for understanding scenes, simulating worlds, and predicting actions. Cosmos 3 targets robotics, autonomous vehicles, and vision systems. Its architecture pairs a reasoning transformer with a generation transformer.

Developers can use the model to interpret a scene, simulate an environment or future state, or build models that generate robot actions. NVIDIA reports training on billions of multimodal samples.

Microsoft introduced seven in-house models at Build. The lineup includes MAI-Thinking-1, its first reasoning model, with 35 billion active parameters and a 256K context window. Access is through Foundry private preview.

Microsoft says it trained the model from scratch without distillation, targeting multi-step instructions, long-context reasoning, and code generation. A separate coding model is available in Copilot and VS Code. The announcement also includes image, voice, and transcription models; transcription supports 43 languages.

Edge added language detection and translation APIs. The browser-native APIs are available in Edge 148 and support more than 145 languages.

Microsoft also previewed Aion-1.0-Instruct in Canary and Dev, including CPU inference on devices without GPUs. Local speech recognition remains a preview too. It uses the Web Speech API so applications can process audio without a cloud connection.

  • A visual guide to Gemma 4 12B (newsletter.maartengrootendorst.com): An illustrated explanation of how removing separate encoders shifts image and audio interpretation into the language model.
  • Code production and operational ownership (charity.wtf): An engineering essay connects increased code production with review capacity, institutional understanding, and operational ownership.
  • Mathematicians ask for AI disclosure and review rules (scientificamerican.com): Reporting on an 11-page declaration describes concerns about the review demands of AI-generated proofs and commercial secrecy’s effects on attribution and openness.

The week in AI, once a week.