Gemma 4's Open Power & OpenAI's Media Move

Google's Gemma 4 brings frontier AI to phones and Raspberry Pi, Alibaba launches Qwen3.6-Plus for agentic coding, OpenAI makes its first media acquisition with tech talk show TBPN, Cursor 3 rebuilds for agents, and Microsoft's MAI models signal growing independence from OpenAI.

Transcript

SPEAKER_00 0:01

Hey, welcome to AI Mornings with Andreas Vig. It's April 3rd, 2026. Google just dropped Gemma 4, and it's a serious upgrade for open source AI. This new model family comes in four sizes, from tiny edge models that run on your phone to a 31 billion parameter powerhouse. The big one, called Gemma 431B Thinking, hits 1452 on Arena AI and solves nearly 90% of AMI 2026 math problems. That's frontier level reasoning in an open model. What's really interesting is the edge optimized versions E2B and E4B. These can run completely offline on phones, Raspberry Pi, and Jets and Nano devices with near zero latency. That means real AI capabilities on tiny hardware without any cloud connection. All four models support 140 languages, handle audio and vision inputs, and have native function calling built in for agentic workflows. The weights are up on Hugging Face, Olama, Kaggle, and LM Studio right now. Speaking of powerful models, Alibaba's Quen team just released Quen 3.6 Plus and they're focusing hard on agentique coding. This thing has a 1 million token context window by default, which is huge for working with large code bases. It sets new state-of-the-art results on Terminal Bench 2.0 at 61.6% and performs really well on SuiteBench Pro. What caught my attention is a new API feature called Preserve Thinking, designed specifically for agents. It keeps the full reasoning trace across multiple turns, which apparently improves decision consistency and can even reduce total token usage. The model is available now through Alibaba Cloud Model Studio. In acquisition news, OpenAI just bought TBPN, the daily tech talk show that's become something of a Silicon Valley institution. This is OpenAI's first move into media ownership. The three-hour live show, hosted by founders John Coogan and Jordi Hayes, is described as Sports Center for Tech, and it pulls in top CEOs like Mark Zuckerberg, Satya Nadella, and Sam Altman himself. The show is reportedly on track to generate more than 30 million US dollars this year. Here's the interesting part. TBPN will report to Chris Lahane, OpenAI's chief political operative. He's the guy who coined vast right-wing conspiracy for the Clinton White House and ran the crypto super PAC Fair Shake. OpenAI says TBPN will keep editorial independence, but an AI company buying a show that covers its competitors right before an IPO raises some eyebrows. Altman posted that he doesn't expect them to go any easier on OpenAI. A few more things worth knowing about today. Cursor just launched version 3, a complete rebuild of their interface around working with AI agents. The new design is multi-workspace, so you can run agents across different repositories at the same time. You can kick off cloud agents from your phone, Slack, or GitHub, then pull them into your desktop for local editing. There's also a new diffs view for reviewing changes and managing PRs without leaving the app. And Microsoft is continuing to build its own AI stack alongside OpenAI. They just announced three MAI foundational models. A transcription model that's 2.5 times faster than Azure Fast, a voice model that generates 60 seconds of audio in one second, and an image generator. The pricing undercuts Google and OpenAI. Microsoft AI CEO Mustafa Suleiman confirmed that their OpenAI partnership isn't going anywhere, but a recent contract renegotiation did unlock their ability to pursue independent superintelligence research.$13 billion invested in OpenAI, and they're still building their own models. That tells you something about where this is all heading. That's it for today. See you tomorrow.