Wednesday, June 3
Wednesday, June 3
Gemma 4 12B Drops the Encoders, Runs Multimodal on Your Laptop
Google DeepMind just released Gemma 4 12B, and the architecture is the interesting part. Traditional multimodal models use separate encoder modules to translate images and audio before the language model processes them. Google skipped that entirely. Gemma 4 12B handles vision and audio natively, no encoder middlemen, and fits in 16GB of VRAM. That's a MacBook. Performance approaches Google's own larger 26B model at less than half the memory. Apache 2.0 licensed, 150 million downloads across the Gemma family, and the local inference community is already racing to quantize it.
Gemma 4 12B Drops the Encoders, Runs Multimodal on Your Laptop
Google DeepMind just released Gemma 4 12B, and the architecture is the interesting part. Traditional multimodal models use separate encoder modules to translate images and audio before the language model processes them. Google skipped that entirely. Gemma 4 12B handles vision and audio natively, no encoder middlemen, and fits in 16GB of VRAM. That's a MacBook. Performance approaches Google's own larger 26B model at less than half the memory. Apache 2.0 licensed, 150 million downloads across the Gemma family, and the local inference community is already racing to quantize it.
Microsoft Build 2026
Microsoft Build wrapped its keynote in Seattle yesterday, and the sheer volume of announcements is still settling. A few things worth knowing going in:
- Build has run since 2011, but this year felt less like a developer conference and more like a product philosophy manifesto. The throughline: apps as we know them might be on borrowed time.
- Mustafa Suleyman, who co-founded DeepMind before joining Microsoft as AI CEO, is now the public face of the company's AI direction. He's steering hard.
- Microsoft spent years positioning itself as the company that partners with model makers. Yesterday it shipped seven first-party AI models in a single announcement. That era appears to be over.
The firehose was real. Here's what landed.
Microsoft Build wrapped its keynote in Seattle yesterday, and the sheer volume of announcements is still settling. A few things worth knowing going in:
- Build has run since 2011, but this year felt less like a developer conference and more like a product philosophy manifesto. The throughline: apps as we know them might be on borrowed time.
- Mustafa Suleyman, who co-founded DeepMind before joining Microsoft as AI CEO, is now the public face of the company's AI direction. He's steering hard.
- Microsoft spent years positioning itself as the company that partners with model makers. Yesterday it shipped seven first-party AI models in a single announcement. That era appears to be over.
The firehose was real. Here's what landed.
Security Scares
Releases, Policy, and the Model Race
Favorite Featured Stories
Every identity mechanism on the web started with a person at a keyboard. When machines needed access, engineers didn't r...
Last week, Anthropic and OpenAI both responded to the same brute constraint: regulated enterprise data cannot leave the ...
A developer's coding agent worked perfectly for thirty minutes, then started contradicting its own decisions. The prompt...
A request times out. You send it again. Retry logic is so fundamental to distributed systems that most frameworks ship i...
Every identity mechanism on the web started with a person at a keyboard. When machines needed access, engineers didn't r...
Last week, Anthropic and OpenAI both responded to the same brute constraint: regulated enterprise data cannot leave the ...
A developer's coding agent worked perfectly for thirty minutes, then started contradicting its own decisions. The prompt...
A request times out. You send it again. Retry logic is so fundamental to distributed systems that most frameworks ship i...