🚀 RELEASES & LAUNCHES
Google finally releases Gemini 4 Argon
After months sitting on the sidelines of the frontier AI race, Google has finally launched its long-awaited Gemini 4 Argon model, claiming it can compete with the best the industry has to offer. Google says Argon delivers "frontier reasoning and improved performance across complex workflows," while
📎 Coverage: theverge.com · 👁 via The Verge AI, Ars Technica AIGoogle Play now hosts private vibe-coded apps built in AI Studio
Google announced a US feature letting users install personal AI Studio vibe-coded apps directly through the Play Store without a computer.- Aimed at casual builders and Android users who create personal apps with Google AI Studio's vibe coding tools.
- The feature launches Tuesday in the US, letting users publish private apps to the Play Store instead of sideloading via USB.
- Google plans to help turn these personal AI-generated apps into wide public releases.
- It removes the previous friction of connecting a phone to a computer to install vibe-coded apps.
- It matters because it shortens the path from prompt to working Android app for non-developers.
Liquid AI releases LFM2.5-24B-A2B: 24B MoE reasoning model for on-device agents
Liquid AI released LFM2.5-24B-A2B, a 24B-parameter MoE reasoning model with only ~2B active parameters, built for fast, private on-device agents.- Aimed at developers building agentic, tool-calling applications that run locally on laptops and edge hardware.
- The model has 24B total parameters with only ~2B active per token, always reasons before answering, and is the largest and final model in the LFM2.5 family (spanning LFM2.5-230M to 24B).
- Scores 48.09 on LiveClaw agentic/tool-use benchmark versus 38.19 for Qwen3.6-35B-A3B, a model with ~50% more total and active parameters.
- Runs up to 1.7× faster than comparable MoEs on-device, hitting 184 tok/s on an Apple M5 Max with up to 21K output tokens; GGUF weights and a free playground demo are available.
- Cheap per-token reasoning plus strong agentic performance makes large-model capability feasible for private, on-device agent deployments.
80B Qwen3-Next MoE runs in a Chrome tab via llama.cpp WebGPU with experts streamed from disk
A developer demonstrated llama.cpp's WebGPU backend running an 80B MoE model in a Chrome browser tab by streaming expert weights from disk.- Targets local-LLM users running MoE models in the browser, with no per-model engine needed — the disk tier sits under llama.cpp's own WebGPU backend.
- Qwen3-Next 80B-A3B Q4_K_M (48.4 GB file, 4.89 GB on GPU) loaded from disk in 48s and decoded at 2.9 tok/s on a MacBook Pro M4 Pro (24 GB) as a proof of concept.
- Smaller models run faster: OLMoE 1B-7B hit 22.4 tok/s (4.2 GB file, 5.3s load) and LFM2 8B-A1B hit 17.1 tok/s (5.0 GB file, 6.2s load); Qwen3-30B-A3B (14.7 GB) also runs.
- The GGUF downloads once into the browser's private OPFS storage, and only active expert weights are paged in from disk per token.
- It matters because it brings disk-streamed MoE inference to any model llama.cpp already supports, entirely client-side in a browser tab.
Facebook brings AI-powered 'Your Algorithm' feed controls to main feed
Meta launched its AI-powered 'Your Algorithm' feature on Facebook's main feed, letting users customize topics they see.- Aimed at Facebook users, the feature lets people add and manage topics they want to see in their feed and Facebook Reels.
- Users can hit 'interested' or 'not interested' on topics, and Meta says recommendations respond right away to updates.
- 'Your Algorithm' first launched on Instagram before coming to Facebook's main feed, per Adam Mosseri's announcement.
- It gives users direct control over AI-driven feed ranking, a shift from purely algorithmic curation.
Tencent open-sources TencentDB Agent Memory, a layered memory engine for AI agents
Tencent Cloud open-sourced TencentDB Agent Memory, an MIT-licensed layered memory engine for AI agents that cuts token use and boosts task success in benchmarks.- For AI agent developers and teams using tools like Claude Code, OpenCode, Codex, CodeBuddy, DeepSeek Harness, Hermes, and OpenClaw.
- The MIT-licensed engine turns conversations, docs, and code into four reusable memory assets: Chat Memory, Skill, LLM-Wiki, and Code-Graph, organized in layers L0 (raw conversation) to L3 (persona).
- Benchmarks as an OpenClaw plugin show up to 61.38% token savings (WideSearch), a 51.52% relative success-rate gain, and PersonaMem accuracy rising from 48% to 76%.
- It runs 100% locally on SQLite (with an experimental MongoDB backend), deploys via one start-all.sh command, and includes a memory hub panel at localhost:8125.
- A Team Memory update extends long-term memory from individual use to team collaboration; the repo has reached ~28.8k GitHub stars.
- It matters because it gives agent builders a free, self-hosted alternative to memory startups, reducing repetitive context re-explanation and token costs.
💰 BUSINESS & FUNDING
Pluto Security emerges from stealth with $42M to secure enterprise AI use
Pluto Security came out of stealth with $42 million in funding for a platform that lets security teams monitor and control how employees use and build with AI tools.- Aimed at CISOs and enterprise security teams; customers include Oscar Health, Workato, Sixt, and AlphaSense.
- Raised $42M total: an $8.5M seed led by TLV Partners (with Armis' founders investing) and a $34M Series A led by Jump Capital, with Skinos and Ansa Capital participating.
- Platform monitors AI tools, skills, extensions, MCP servers, and employee-built agents, and can intervene mid-session — stopping risky actions or injecting instructions so agents ask for confirmation.
- Traction since starting sales in April 2026: $2M+ ARR, 30+ enterprise customers, and 1.5M+ endpoints secured; ~60% of detected risks are in tools vibe-coded by non-developers, led by finance teams.
- Free Plutonium risk catalog and MCP scanner draws thousands of security pros, feeding an inbound pipeline that now makes up 70% of sales; new capital goes to product, research, and US expansion.
SpaceXSI bundles X Premium, Grok, and Cursor into one X Subscription
SpaceXSI launched The X Subscription, bundling X Premium, Grok, and Cursor into plans from $8 to $200 per month.- Aimed at X users, AI chat users, and developers, the bundle combines Grok chat/voice/images/video, the agentic Grok Bot, Cursor AI coding, and X Premium perks like the verified checkmark and fewer ads.
- Five tiers: Free, Plus ($8/mo), Premium ($30/mo), Super ($100/mo), and Ultra ($200/mo), with usage pools 4x, 15x, and 50x Plus at Premium, Super, and Ultra respectively.
- Premium or higher is required for the latest Grok models; third-party models in Cursor are billed separately at API list prices.
- Existing X Premium+ subscribers can switch to a comparable plan at the same or lower price, with automatic refunds for unused time; upgrades roll out gradually.
- The bundle positions SpaceXSI against OpenAI, Anthropic, and other paid AI subscription rivals.