join.dev
AI news · 10 stories
Google launches Gemini 4 Hydrogen for fast agent workflows · IBM previews LLMON models for explicitly structured prompts · AWS releases an agent-built emulator for local cloud testing
Google introduced Gemini 4 Hydrogen on November 12 and began rolling it out across developer, enterprise and consumer products. The model targets fast coding and agent execution, while a new Antigravity Orchestrator switches models within a session.
Read at Google
IBM released experimental Granite 4.1 3B and 8B previews on November 12 that use LLMON to distinguish instructions, data and their relationships. The accompanying toolkit supports evaluation of this approach to prompt structure; the models are not intended for production.
Read at IBM
AWS launched AWS Emulator on November 11, providing a free, offline environment for testing applications and coding agents against AWS-style APIs. AWS says agents wrote and tested the code, with fidelity tests and automated review gates checking behavior against real services.
Read at All Things Distributed
Runway released Praxis-1 weights and inference and fine-tuning code under Apache 2.0 on November 11, following its earlier partner rollout. The model predicts video and actions together, giving robotics developers a starting point they can adapt to their own hardware.
Read at Runway
Google introduced Gemini Decision 1 and the Decisions API on November 11 for scoring bounded choices across text, images, audio, video and PDFs. A single request can evaluate up to 1,000 questions and return confidence scores, supporting tasks such as tool routing and policy checks.
Read at Google
Zyphra published ZONOS3-Mini with Apache 2.0 weights for multilingual speech synthesis, voice cloning and delivery control. The November 12 release post describes 100 million active parameters and real-time on-device execution, expanding options for locally run voice applications.
Read at Zyphra
Baidu released ERNIE 5.5 Flash Preview as a post-trained model built on its earlier Flash Base, with API access through Qianfan. It targets coding, 3D design and knowledge work; training is ongoing and Apache 2.0 weights are promised rather than already released.
Read at Baidu
Hugging Face published a November 12 guide with scripts and containers for adapting TRELLIS.2 to custom 3D assets locally or on Hugging Face Jobs. Its LoRA workflow focuses on geometry and structure, giving developers a reproducible way to specialize 3D generation.
Read at Hugging Face
Four nonprofits filed a November 12 lawsuit in San Francisco seeking to freeze advanced AI development at Anthropic, Google, Meta, OpenAI and SpaceXSI. The complaint invokes California unfair-competition law and alleges serious safety risks, putting those claims before a court.
Read at Public Citizen
The IMF said on November 12 that it was monitoring AI equity volatility and political changes that could slow data-center development. It said it was too early to conclude that investors were making a broader, lasting reassessment of AI profitability and productivity gains.
Read at Reuters