Executive Summary
- Google DeepMind has appointed Koray Kavukcuoglu as its new chief, reporting directly to Sundar Pichai, as the Gemini application achieves one billion monthly active users.
- Meta has released Muse Glimmer, an open-weight 30B multimodal model under an Apache 2.0 licence, optimised for local agentic workflows within 24GB–32GB RAM envelopes.
- OpenAI has expanded ChatGPT, Work, and Codex to Linux desktops, whilst its backed enterprise vehicle Thrive Holdings raised $2bn at a $12bn valuation.
- Anthropic has introduced global invisible text watermarking and C2PA file metadata for Claude outputs, aligning directly with EU Code of Practice guidelines.
- NVIDIA has launched Nemotron 3.5 Lightning, a 30B MoE hybrid combining Mamba-2 and Attention architectures for efficient local agent execution on single-GPU hardware.
In Detail
TechCrunch
- OpenAI launches ChatGPT desktop app for Linux
- Google's Gemini app surges to 1 billion monthly users
- OpenAI-backed Thrive Holdings raises $2B at $12B valuation
- AI coding startup Cognition in talks to raise at $40B valuation
VentureBeat
- Enterprise AI organizations struggle with metering agent compute costs
- Four of five enterprises with secured AI agent identities cannot contain rogue execution
- AI context layers double bad-answer detection in governed enterprises
The Verge
- Grok launches beta cloud-based AI teammate agents
- ChatGPT and Gemini both cross 1 billion active user threshold
Hugging Face
- Alibaba releases open-weights Qwen3.8 (2.4T total, 95B active MoE)
- NVIDIA releases Nemotron-3.5-Lightning-30B-A3B-NVFP4 weights
Ars Technica
- Websites adopt custom font rendering to disrupt AI scrapers
- Twitch introduces user opt-out for Amazon AI training data
Model Releases
- Muse Glimmer — Meta. Meta released Muse Glimmer, an open-weight 30B parameter model released under an Apache 2.0 license designed to execute multi-step agentic workflows on local hardware. Read more
- Nemotron 3.5 Lightning — NVIDIA. NVIDIA introduced Nemotron 3.5 Lightning, a 30B total / 3B active MoE hybrid (Mamba-2 + Attention) supporting context lengths up to 1M tokens for local agent execution. Read more
- Qwen3.8 — Alibaba. Alibaba open-sourced Qwen3.8, featuring 2.4T total parameters with 95B active per token, operating natively in thinking mode. Read more
What to Watch
Interpretation, not reported fact. Each point follows from a source listed below — follow the links and judge for yourself.
- Shift towards localized hardware execution for agentic AI workloads may lower enterprise inference costs. (Basis: Concurrent releases of Meta's Muse Glimmer and NVIDIA's Nemotron 3.5 Lightning designed specifically for single-GPU local agent execution.)
- Enterprise focus will increasingly pivot from basic agent deployment to governance and runtime containment. (Basis: VentureBeat research reporting 80% of enterprises securing AI agent identities remain unable to contain rogue agent behavior.)
Sources
- TechCrunch (https://techcrunch.com)
- VentureBeat (https://venturebeat.com)
- The Verge (https://theverge.com)
- Hugging Face (https://huggingface.co)
- Ars Technica (https://arstechnica.com)