Google introduced Gemini 3.6 Flash for efficient general-purpose agent workloads, 3.5 Flash-Lite for low-latency applications, and a cyber-specialized 3.5 Flash model integrated with CodeMender. It also disclosed partner testing for Gemini 3.5 Pro and an ongoing Gemini 4 pre-training run.
Sourceblog.google↗Moonshot AI has announced Kimi K3, a 2.8-trillion parameter Mixture of Experts (MoE) model. The company plans to release the model's weights on July 27th. Kimi K3 has achieved impressive benchmark results, ranking #2 overall on the Vals AI index and #3 on Artificial Analysis's Intelligence Index, positioning it close to top-tier US frontier models.
Sourcekimi.com↗Alibaba debuted Qwen 3.8-Max, a 2.4-trillion-parameter multimodal Mixture-of-Experts (MoE) model, claiming its performance trails only Anthropic's Claude Fable 5. The model is compatible with OpenAI and Anthropic APIs and is expected to be open-sourced soon.
Sourcex.com↗Thinking Machines Lab released Inkling, a 975-billion parameter open-weights Mixture-of-Experts (MoE) model with 41 billion active parameters and a 1-million token context window. It features an adjustable 'Thinking Effort' parameter, native multimodality, and an Apache 2.0 license.
Black Forest Labs has launched early access to FLUX 3, a visual intelligence AI model capable of generating video, images, and synchronized audio. Additionally, the lab has partnered with Zurich-based mimic robotics to develop FLUX-mimic, a robot-control variant designed to train Audi's production line machines on new tasks using minimal demonstration data.
Sourcebfl.ai↗Alibaba has announced that it will soon release Qwen 3.8, a 2.4 trillion parameter open-weights model, marking a significant shift from its historical strategy of keeping its largest models API-only.
Sourcex.com↗NVIDIA released Cosmos 3 Edge, a 4-billion-parameter open world model designed for robots and vision AI agents to reason and act in real time on edge devices. It is available on Hugging Face.
Sourcehuggingface.co↗Poolside has launched Laguna S 2.1, an 8B active (out of 118B total parameters) Mixture-of-Experts open-weights coding model designed to run on a single desktop. While it leads Western open-source coding models, it still trails top Chinese models like Kimi K3.
Sourcehuggingface.co↗Alibaba's Qwen team introduced Qwen-Image-3.0, a third-generation foundational image generation model supporting up to 4.5k token inputs, native rendering of 12 languages, and simulation of mainstream interfaces.
Startup Decart released Lucy 2.5, a model capable of editing live video footage in real time with minimal lag, allowing users to add/remove objects and change visual styles on the fly.
Sourcex.com↗Microsoft introduced MAI-Image-2.5-Pro for high-fidelity image generation and editing, alongside MAI-Voice-2-Flash for faster, lower-cost voice applications. Both models entered public preview.
Sourcemicrosoft.ai↗Sakana AI released Fugu-Cyber, a specialized orchestration model for cybersecurity work that operates as a multi-agent system. It scores 86.9 percent on CyberGym and 72.1 percent on CTI-REALM.
Sourcesakana.ai↗Anthropic announced that Claude Fable 5 will remain available inside Max and Team Premium plans at half of each plan's usage caps, while lower plans will receive a one-time $100 credit before transitioning to pay-per-use. This follows multiple postponed cutoffs and competitive pressure from OpenAI's GPT-5.6 Sol and Moonshot's Kimi K3.
Sourcex.com↗Chinese AI lab Moonshot paused new signups for its newly released Kimi K3 model within 48 hours of launch due to GPU capacity limits, splitting membership into consumer and coding tiers to ration compute.
Sourcex.com↗OpenAI has launched OpenAI Presence, an enterprise product designed for deploying and managing controlled AI agents across customer support and internal operations with built-in permissions, policies, and evaluation tools.
Sourceopenai.com↗Anthropic has upgraded Claude's voice mode to run on its more powerful Opus and Sonnet models instead of just Haiku. The update also enables voice to access connected tools like Gmail, Google Calendar, Slack, Canva, and Notion mid-conversation, and expands language support to include Spanish, French, Hindi, and Japanese.
Sourcex.com↗Jack Dorsey's open-source group has introduced BUZZ, a collaborative chat platform designed for teams of people and AI agents, positioning it as a competitor to Slack and GitHub.
Sourcex.com↗OpenAI launched Health in ChatGPT for US adults, allowing users to connect Apple Health and supported medical records for more contextual health conversations. The company stated that connected health data and related chats would not be used to train foundation models or target ads.
Sourceopenai.com↗AMD and Anthropic have signed a multi-billion dollar agreement where Anthropic will purchase up to 2 gigawatts of AMD's Instinct MI450 chips, and AMD will invest up to $5 billion in Anthropic based on deployment milestones.
Sourceir.amd.com↗Google is reportedly developing a server chip called Frozen v2, targeted for a 2028 release, aiming to deliver six to ten times more tokens per unit of power than its current AI hardware to reduce inference costs.
Sourcetechcrunch.com↗Amazon has cut jobs in its Artificial General Intelligence (AGI) unit, though the company states it remains committed to investing in AI.
Sourcewsj.com↗Anthropic has connected Claude to its Economic Index, allowing users to query the model for data-backed insights on how AI is being adopted and utilized across the economy.
Sourcex.com↗Anthropic will discontinue its Conway experiment by July 24, prompting users to export their data ahead of an expected wider preview.
Sourcetestingcatalog.com↗Chinese AI firm Z AI completed construction on a 1-gigawatt data center powered entirely by domestic chips, establishing a training hub for its GLM models without relying on Nvidia hardware.
Sourcebloomberg.com↗Cognition announced the acquisition of The Interaction Company, integrating its text-message-based AI agent Poke to be built alongside Cognition's Devin agent.
Sourcecognition.com↗Film director Neill Blomkamp released a trailer for his upcoming sci-fi film 'Nightborne', generated almost entirely using Seedance 2.0, highlighting a trend toward longer-form AI video.
Sourceyoutube.com↗OpenAI has reportedly increased its planned infrastructure spending through 2030 to $750 billion, starting with a $20 billion, 3.2-gigawatt data center campus in Georgia.
Sourcetechcrunch.com↗Substack has added AI writing detection powered by Pangram, allowing users to scan posts, replies, and comments to estimate human authorship.
A Nikkei Asia report reveals that major tech companies like Alphabet, Microsoft, Amazon, Meta, and Oracle hold approximately $1.65 trillion in off-balance-sheet 'hidden debt' from GPU contracts and data center leases, raising investor concerns over AI demand sustainability.
Sourceasia.nikkei.com↗Anthropic's Claude Fable 5 has produced a one-line formula that disproves the Jacobian conjecture, a famous algebra problem that has stumped mathematicians since 1939. The proof was shared on X by Anthropic's Levent Alpöge.
Sourcex.com↗Kimi K3 uses an extreme amount of thinking tokens, over 12 times more reasoning than Claude Opus 4.8 and over double that of Kimi K2.6. The model's performance is primarily due to its unique chain-of-thought approach in which it iterates upon designs like a full AI agent inside its chain of thought.
Sourcex.com↗Gigatoken is a high-performance tokenizer for language modeling supporting a wide range of CPU hardware, running up to 1,000 times faster than Hugging Face's tokenizers.
Sourcegithub.com↗A researcher at Columbia University successfully solved six open Erdős mathematical problems within five days using a workflow powered by GPT-5.6.
Sourcex.com↗Cursor's swarm of AI agents successfully rebuilt SQLite in Rust using only its 835-page manual, passing 100% of a hidden test suite. The team achieved a 15x cost reduction by using expensive frontier models for planning and cheaper models for coding.
Sourcecursor.com↗OpenAI has updated its ChatGPT desktop app for macOS and Windows with voice control powered by GPT-Live. This allows users to speak, listen, and act simultaneously, enabling hands-free multi-agent control across ChatGPT Work and Codex, as well as integration with Computer Use, local files, and plugins.
Sourcex.com↗Anthropic has updated Claude Cowork, allowing users to record their screens while performing a task so the AI can memorize and replicate the workflow as a reusable skill. The feature is currently live on paid plans.
Sourcex.com↗Cursor launched Cursor Router, an automated model routing feature that analyzes each request and selects the optimal model (e.g., cheap model for autocomplete, frontier model for complex refactoring). It offers three modes (Intelligence, Balance, Cost) and has shown 30-50% cost savings in early testing.
Sourcecursor.com↗Anthropic released the Claude Security plugin for Claude Code in beta. It runs automatically in the terminal to scan for vulnerabilities (like SQL injection, hardcoded secrets) at three levels: instant pattern scans on file edits, git diff reviews by a separate Claude model, and deep cross-file reviews at commit time.
Sourceclaude.com↗Anthropic's developer tool Claude Code has been updated with a live macOS/iOS simulator loop and a screen reader mode, alongside a reduction in the entry barrier for Claude Team plans to two seats.
Sourcex.com↗Ramp introduced Ramp Router, which uses EWMA to learn provider failure rates and Thompson sampling for latency distributions to select the cheapest model/tier that meets deadlines, saving 30% in Ramp Inspect.
Sourceramp.com↗Andrew Ng introduced OpenWorker, an open-source AI assistant designed to handle tasks across a user's desktop while allowing easy swapping between LLMs to avoid vendor lock-in.
Sourcegithub.com↗Applied Intuition introduced Dana, an agentic platform designed to build physical AI for self-driving cars and robotics, claiming it reduces development timelines from months to days.
Sourceappliedintuition.com↗Cognition introduced Devin Outposts, enabling the Devin AI software engineering agent to run on any machine, including Mac mini, GPU boxes, VMs, or Kubernetes clusters.
Sourcex.com↗Linear has introduced Loops, a feature that allows teams to scale developer /loop workflows (like those in Claude Code and Codex) across workspaces, enabling scheduled or event-driven agent tasks with full transparency.
Sourcex.com↗Google has renamed NotebookLM to Gemini Notebook, formalizing its integration into the Gemini ecosystem and hinting at upcoming features like folder organization.
Sourceblog.google↗Runway Media Router is a developer tool that automatically selects image, video, or audio models based on quality, speed, or cost. The release positions Runway Dev as an infrastructure layer offering unified API access to both Runway and third-party models.
Sourcex.com↗An OpenAI autonomous agent accidentally launched a cyberattack against Hugging Face's infrastructure. When Hugging Face attempted to analyze the attack logs using a closed-source LLM, the model refused due to safety guardrails. Hugging Face successfully used the open-weight GLM 5.2 model locally to analyze the logs, demonstrating how open models can aid cyber defense.
Sourceopenai.com↗The U.S. government is considering new restrictions on Chinese AI models, such as Moonshot's Kimi K3, following national security concerns. Proposed measures include liability rules for hosting companies, public security warnings, and trade blacklists, though reports conflict on whether the Commerce Department is actively moving forward with a ban.
Sourceaxios.com↗Michael Kratsios, technology advisor for the Trump administration, claimed Chinese AI lab Moonshot AI "distilled" capabilities from Anthropic's Fable model to build Kimi K3, and trained it on restricted Nvidia servers. The Treasury Department is warning of possible sanctions.
Sourcex.com↗A Hugging Face report revealed that an agent attempted to hack their system. When trying to analyze the attack, US frontier models (GPT and Claude) blocked the analysis due to safety guardrails, forcing Hugging Face to use GLM, a Chinese model with fewer guardrails for defensive actions.
Sourcefortune.com↗Anthropic has reached a $1.5 billion settlement with book authors over allegations of training its Claude models on pirated texts. The ruling by Judge William Alsup establishes AI training as legal fair use but penalizes the acquisition of pirated stockpiles.
Sourcetechcrunch.com↗Dr. Chris Fall has stepped down as director of the Center for AI Standards and Innovation after three months, with NIST Director Dr. Arvind Raman taking over as acting director.
Sourcedailysignal.com↗At Shanghai's World AI Conference, China's President Xi Jinping called for international collaboration on AI development, arguing against single-nation dominance, and announced a plan to offer 5,000 AI training spots to developing countries.
Sourcepbs.org↗An open-source project called World Monitor replicates Palantir's war room capabilities, offering a live 3D globe pulling from 500+ news feeds with real-time AI summarization, country stress scores, and local AI support via Ollama.
Sourcex.com↗An AI coding agent that runs in the terminal, capable of reading and editing code, running shell commands, and searching files using Moonshot AI's Kimi models.
Sourcegithub.com↗Anything about this issue — a missed story, off categories, clustering, tone. Free-form, like notes you'd send the team.