Latest AI news
Curated AI news on models, agents, tools, research and industry, in English, Spanish and Brazilian Portuguese.
ModelsGoogle Search rolls out Gemini 4 Hydrogen, its largest model upgrade so farGoogle has made its Gemini 4 Hydrogen model available globally to all Search users, promising better intent understanding, improved query fan-out, and more visual responses.ResearchApple Study Characterizes the Theoretical Limits of Language Model AlignmentA paper by Apple researchers derives information-theoretic bounds on reward improvement under a KL-divergence budget, showing best-of-N sampling nearly reaches the limit while PPO and GRPO remain suboptimal.AgentsAWS unveils next-generation AgentCore runtime with faster starts, larger sessions and data-level isolationAmazon's updated Bedrock AgentCore runtime starts agent sessions up to 15x faster, scales to 8 vCPUs and 32 GB per session, extends the isolation boundary to mounted storage and adds usage-based billing.
ModelsGoogle introduces Gemini 4 Hydrogen, a fast model for everyday and agentic workflowsGoogle launches Gemini 4 Hydrogen, a speed-focused model for coding and agentic work, across the Gemini app, Search, AI Studio, the Gemini API, and enterprise surfaces, with flat pricing over a 1 million-token context.ToolsLessonFolk offers free, open-source AI lessons through an AI tutor in your existing chat appLessonFolk is an open-source learning platform that teaches AI through conversational lessons delivered by an AI tutor over MCP, inside the chat application users already have.IndustryInception Labs Winds Down Mercury API for Its Diffusion ModelsIn an email reproduced on Hacker News, Inception Labs said the Mercury API hosting its diffusion language models stays fully available through 28.12.2026; a Discord note adds removal from OpenRouter from 08.11.2026.ToolsSPECTRA compresses LLMs roughly 3x on CPU alone, with no calibration stepNeuraCryption published technical documentation for SPECTRA, an engine that shrinks language models between 2.80x and 3.31x without a GPU or calibration datasets. All reported figures are self-reported by the company.AgentsKrea Opens Beta for Krea Bots: Nine Named AI Agents for Creative CampaignsKrea opened a beta for Krea Bots, nine named AI agents that run creative production inside a shared team chat, with $1,000 in credits for accepted teams and pricing still undisclosed.AgentsHark rolls out Handoff², a second-generation browser agent it claims is twice as fastHark began deploying Handoff², the new version of the agent behind its Hark Pro assistant, reporting 82% completion on an internal benchmark, 46% less time per task, and a result within five points of Claude Opus 5.5.AgentsFish Audio Launches Fish Agent Beta on Its Live 3 Full-Duplex Speech ModelFish Audio's public-beta Fish Agent platform lets developers build phone-capable AI voice agents in about 30 seconds on the new Live 3 duplex model, priced at $0.035 per minute during the launch promotion.
IndustryGoogle Cloud extends partner Cross-Cloud Interconnect to last-mile connectivity with LumenGoogle Cloud has launched a preview of partner Cross-Cloud Interconnect for last mile, using Lumen Technologies' network to deliver dedicated Layer 3 private connectivity to private clouds, on-premises sites, and neoclouds.IndustrySentinelOne and Google Security Operations unveil re-architected telemetry integration for agentic SOC workflowsGoogle Cloud and SentinelOne have rebuilt their integration to stream endpoint telemetry directly into Google Security Operations via APIs, cutting ingest latency, storage costs, and event sizes for detection and agentic SOC pipelines.