AI News Digest — April 3, 2026
A landmark day for open-weight models. Google dropped the Gemma 4 family with permissive Apache 2.0 licensing, Microsoft unveiled a trio of foundational models under the MAI banner, and ElevenLabs pivoted from voices to full music generation. Here’s everything you need to know.
Google Gemma 4 — Apache 2.0 Vision-Capable Reasoning Models
Google released the Gemma 4 family, and this is the biggest open-model drop of the week. Four models, all Apache 2.0 licensed, all vision-capable, all designed for reasoning tasks:
- Gemma 4 E2B — 2 billion effective parameters
- Gemma 4 E4B — 4 billion effective parameters
- Gemma 4 31B — 31 billion parameters
- Gemma 4 26B-A4B — Mixture-of-Experts with 26B total, 4B active parameters
The E2B and E4B variants also support audio input, making them true multimodal models at the small end of the spectrum. All four introduce Per-Layer Embeddings, a new technique that improves how the models process and route information through their transformer layers.
From a developer standpoint, the Apache 2.0 license is the real headline. Unlike some “open” models with restrictive commercial terms, Gemma 4 gives you full freedom to use, modify, and deploy commercially. The smaller models (E2B, E4B) are ideal for edge deployment, on-device inference, and fine-tuning on consumer hardware.
Simon Willison already has GGUF quantized versions running in LM Studio and has released llm-gemini 0.30 with support for the new models, so you can query them directly from the command line via the Gemini API.
Microsoft MAI — Three New Foundational Models
Microsoft launched three models under the MAI (Microsoft AI) brand, available on Microsoft Foundry:
- MAI-Transcribe — A speech-to-text foundational model targeting high-accuracy transcription across multiple languages and acoustic conditions
- MAI-Voice — A voice generation model for natural-sounding speech synthesis
- MAI-Image-2 — The second generation of Microsoft’s image generation foundational model
The MAI series signals Microsoft’s push to own the foundational layer across modalities — not just text, but audio and visual generation too. All three are positioned as building blocks for developers integrating AI into enterprise applications. If you’re building on Azure, these integrate directly with the existing Azure AI Services ecosystem.
ElevenLabs Enters AI Music Generation
ElevenLabs, best known for AI voice cloning and text-to-speech, launched a music generation app. The move expands the company beyond spoken audio into full music composition and production.
Details are still emerging, but the app reportedly lets users generate complete musical tracks from text prompts — similar to what Suno and Udio offer, but leveraging ElevenLabs’ audio expertise. For developers building creative tools, this adds another API-powered option for programmatic audio generation.
GitHub Copilot: GPT-5.4 Mini and Data Usage Policy Changes
Two important GitHub Copilot updates landed today:
GPT-5.4 mini in Copilot Student: OpenAI’s compact GPT-5.4 mini model is now available in the auto model selection for Copilot Student tier. This gives students access to a modern, capable model without the compute overhead of the full GPT-5.4.
Data usage policy change (April 24): Starting April 24, GitHub will begin using Copilot interaction data for model training unless organizations explicitly opt out. If you’re using Copilot in a corporate environment, you need to review your settings and opt out if your organization’s data governance policy requires it. This is a significant shift — previously, Copilot telemetry was not used for training by default.
OpenAI Acquires TBPN
OpenAI acquired TBPN, the buzzy founder-led business talk show. The acquisition is a surprising one — TBPN is a media property, not a technology company. The move suggests OpenAI is investing in content creation and distribution channels, potentially to build out a media ecosystem around its AI products.
Granola Note-Taking App: Privacy Red Flags
Privacy researchers flagged significant concerns with Granola, the AI-powered note-taking app that’s been gaining traction:
- AI training is enabled by default — your meeting notes and transcripts are used to train Granola’s models unless you dig into settings to disable it
- Link-sharing exposes notes — anyone with a shared link can view your notes, with no authentication gate
If you or your team uses Granola for meeting notes, check your settings immediately. Disable AI training if you’re handling sensitive or proprietary information, and audit any shared links.
Holo3: Open Computer Use Agent
Hcompany released Holo3, an open computer use agent that can autonomously navigate desktop applications, click buttons, fill forms, and complete multi-step workflows. Think of it as an open-source alternative to Anthropic’s computer use capability.
The agent uses visual understanding to interact with GUIs the way a human would — no API integration required. For developers building automation tools, testing frameworks, or accessibility features, Holo3 is worth a close look.
Anthropic “Mythos” — Next Model Name Leaked
Anthropic’s next model has a name: Mythos. The moniker was revealed through a security lapse that exposed internal references. No technical details or release timeline yet, but the leak confirms Anthropic is actively developing its next-generation model.
Simon Willison’s Corner: Gemma 4, Pretext, and Vulnerability Tooling
Simon Willison continues to ship useful tools and provide incisive commentary:
llm-gemini 0.30: Updated his llm CLI tool to support Gemma 4 models (gemma-4-26b-a4b-it and gemma-4-31b-it) via the Gemini API. Available now via pip install llm-gemini.
Pretext: A new browser library by Cheng Lou (ex-React core team) for calculating paragraph height without touching the DOM. This is a surprisingly hard problem — you need to know how tall text will be before you render it. Pretext solves it efficiently, which is useful for virtualized lists, text truncation, and layout calculations.
Python Vulnerability Lookup: Willison built a tool that checks your pyproject.toml dependencies against the OSV.dev vulnerability database. Run it against your projects to catch known CVEs.
Mr. Chatterbox: Trip Venturella’s Victorian-era ethically trained LLM got a writeup — a quirky but interesting experiment in training models on curated, ethically sourced historical text.
Quick Hits
- Cognichip raised $60M for AI-powered chip design, aiming to accelerate semiconductor development with generative AI.
- Baidu robotaxis froze en masse in traffic, creating chaos on Chinese streets. A reminder that autonomous vehicle reliability still has edges to work out.
- Okta CEO announced a major bet on AI agent identity — building authentication and authorization infrastructure specifically for AI agents acting on behalf of humans.
- Simon Willison got a 128GB M5 MacBook Pro and has been vibe-coding SwiftUI experiments, reporting that the extra RAM makes local LLM inference significantly more practical.
Tool & Release Radar
| What | Details |
|---|---|
| Google Gemma 4 (E2B, E4B, 31B, 26B-A4B) | Apache 2.0 vision-capable reasoning models |
| Microsoft MAI-Transcribe, MAI-Voice, MAI-Image-2 | Foundational models on Microsoft Foundry |
| ElevenLabs Music App | AI-powered music generation from text prompts |
| llm-gemini 0.30 | CLI support for Gemma 4 models via Gemini API |
| Pretext | DOM-free paragraph height calculation for browsers |
| Holo3 | Open-source computer use agent by Hcompany |
| GPT-5.4 mini | Now in GitHub Copilot Student auto model selection |
This digest is published daily at 5pm AEST. Got a tip or a tool we should cover? Drop it in the comments. Tomorrow’s edition will cover developments from April 4, 2026.