The gist: AGIBOT unveils GO-3 Preview and G2 Max, Audi tests Agile Robots' humanoids in Germany, Rokae ships its 10,000th force-controlled arm, plus three funding rounds.
The gist: At Apsara Amsterdam, Alibaba Cloud CTO Feifei Li said Qwen4 ships next week; a slide listed Max, Plus, Flash and 27B. Developers ask if 27B fits a 3090/5090.
The gist: Beyond Fusion, GitHub Universe shipped Automations, goal-based Projects, Dreaming for memories, and a Security canvas in the Copilot app.
The gist: BotGauge is a trust layer for AI agents: adaptive red teaming, full traces, and failures turned into regression evals. Now on Product Hunt.
The gist: Rob LoCascio's Uare.ai raised $10.3M led by Mayfield and Boldstart for Individual AI and its Human Life Model—beyond a generic chat box.
The gist: Cartesia's updated H-Nets learn chunks from raw bytes, not BPE. A 2-stage model reportedly matched a ~3B open LLM with about 10× less data and compute.
The gist: Suhail Doshi's Particle launches Particle-1 on DeepSeek V4.1 Flash. It claims ~45% fewer output tokens on coding agents, with accuracy close to the base.
The gist: Mistral's Responses API is in public preview for agent builders, returning answers, reasoning and tool calls as separate items, in OpenAI Responses format.
The gist: Fireworks' Specialized Intelligence Index adds real-work evals from 31 partners incl. Zapier, Upwork, Grafana and Hebbia, comparing models by domain.
The gist: GitHub open-sourced Replica (MIT). It builds interactive prototypes in the Copilot app from a team's own components, design guidance and docs.
The gist: Paper Instruments' model for finance, law and consulting is open-weight on Fireworks and Baseten. The company says it beats Claude Fable 5 on its own benchmark.
The gist: Rime's Prism is an end-to-end speech-to-speech model for enterprise voice agents. It writes replies as text first, and the company says it can be tuned with RL.
The gist: Vercel's open-source coding agent fx is now a built-in T3 Code provider on web, desktop and mobile, with Supervised and Auto modes and your existing models.
The gist: Meta is giving researchers at Caltech, Ohio State's College of Medicine and three other US universities its latest Muse models, with $2 million in grants.
The gist: Cognition's FrontierCode 2.0 data reportedly listed an unranked Obsidian 3 at 46.8%, above Opus 5.5's 45.5%, then dropped it. Anthropic has said nothing.
The gist: A developer compiled a 32-bit RISC-V CPU into Qwen3.5-4B's weights, using its Gated DeltaNet layers as RAM to run C and MicroPython with no tools. Open source.
The gist: The Gemini app now offers interactive, visual explorations that unpack topics layer by layer, from deep space to jet engines. It starts in the student hub.
The gist: OpenAI's always-on dot agent can now be texted via iMessage and RCS with text, photos or voice, and it texts back proactively. US ChatGPT Pro preview only.
The gist: Dalet Labs has launched Scene, an AI shopping app that generates realistic scenes around your tastes, with real clothing and home products you can buy in the app. The company also announced a $7 million pre-seed round led by Lightspeed. Scen
The gist: GitHub Copilot Fusion is generally available. It picks and combines models from Anthropic, OpenAI and others for each task, in Standard, High and Max tiers. In an Artificial Analysis evaluation cited by GitHub, Max trails Claude Code with Opus 5.5 by two points at about half the cost per task. It works on every plan, w
The gist: NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model with about 30B total and 4B active parameters that handles 40-language transcription, speaker attribution and translation in one checkpoint, with prompt-based control over terminology. Weights are on Hugging Face.
The gist: Google DeepMind has unveiled Gemini 4 Carbon, which it says ranks first on the Artificial Analysis Intelligence Index v5 and sets new bests in cybersecurity and long-horizon tasks. It opens this week to trusted defenders through the Fairwind Program before a wider rollout.
The gist: India's Central Institute of Indian Languages has released Anuvani-LID and Anuvani-Diarization, two open speech models for Indian languages. One identifies the spoken language and the other separates speakers, covering the 22 scheduled languages plus Indian English.
The gist: Cisco has open-sourced Project Quarry, a detection rule language and engine that spots risky multi-step AI agent behavior across a whole session, such as reading credentials and then sending data out. Rules look like Snort or YARA, and plugins for Claude Code and Codex are included.
The gist: SenseTime has released SenseNova 6.8 Code2Video, a 27B open model under Apache 2.0 built on Qwen3.8-27B. Instead of generating pixels, it writes executable code and renders it into short motion-graphics and explainer videos.
The gist: CopilotKit has released OpenDuet, an open-source, self-hostable template for a proactive AI agent and an open alternative to Claude Duet. It works with any agent framework, asks for approval before sensitive actions, and runs on web, mobile and chat apps.
The gist: Google’s Gemini Robotics Embodied Reasoning 2 is in public preview on Google Cloud / Model Garden. As a robot’s high-level brain, it reasons over video for spatial and progress cues, plans multi-step work, and can call low-level control and navigation APIs across embodiments from arms to humanoids.
The gist: Meta’s personal AI agent Muse is now available in Singapore, with connectors for Gmail, Calendar, and Instagram plus local services such as Grab, Shopee, and Singapore Airlines. It is free with a weekly usage cap and was already live in the US, Canada, and Japan.
The gist: Axiom Math used AxiomProver AI-agent swarms to produce a complete Lean formalization of CFSG in about 24 days—roughly 43.1 million lines of code—and flagged multiple deviations in the published literature.
The gist: Uncle Bob open-sourced ub-mode: an AI coding-agent skill that runs first, tests first, with complexity, duplication and mutation-test gates.
