The whole family ships under Apache 2.0, which means you can actually build a business on top of it without lawyers breathing down your neck. Four variants from 2B parameters up to 31B dense, with context windows up to 256K tokens and native vision and audio. The smallest ones run on a Raspberry Pi. I'm not exaggerating. Google is basically handing out what would have been a frontier model two years ago for free.
If you're still paying per-token for tasks that a local model can handle, this is your wake-up call. The cost floor for AI just dropped again.
OpenAI Eyes an IPO as Revenue Hits $25 Billion
OpenAI crossed $25 billion in annualized revenue and the rumor mill says they're taking early steps toward going public, maybe as soon as late 2026. Anthropic isn't far behind at $19 billion. These are real businesses now, not research labs burning through venture capital.
An IPO would change the dynamics in ways that matter to builders. Public companies answer to shareholders, and shareholders want predictable revenue. That usually means more enterprise features, more stable APIs, and less "we're going to pivot the product on you overnight." It also means quarterly earnings calls where we get actual numbers instead of leaked memos.
My take: this is mostly good news for developers. Competition at this scale keeps prices falling and quality rising. Just don't be surprised when the "for the benefit of humanity" messaging gets a lot quieter once there's a stock ticker involved.
MCP Crosses 97 Million Installs
Anthropic's Model Context Protocol hit 97 million installs in March. Every major AI provider now ships MCP-compatible tooling. That's not a protocol anymore. That's a standard.
I've been building with MCP for months and the ecosystem growth is wild. Six months ago you had to wire up everything yourself. Now there are off-the-shelf connectors for just about every SaaS tool your company uses. The protocol basically solved the "how do I give my AI agent access to my stuff" problem that everyone was hand-rolling custom solutions for.
What matters here isn't the install count. It's that we finally have a common interface between AI models and the tools they need to use. That's the kind of boring infrastructure win that makes entire categories of applications possible. If you're building anything agentic and you haven't looked at MCP yet, you're doing it the hard way.
Federal Judge Blocks Government AI Ban
A federal judge ruled that the Trump administration violated free speech protections by banning certain AI models from government systems. The specifics of the ruling are still shaking out, but the bigger picture is clear: the government can't just pick winners and losers in AI by executive fiat.
This matters because government contracts are a massive market, and arbitrary bans based on politics rather than capability create real distortions. Federal agencies should be choosing AI tools based on what works best for their mission, full stop.
I'm not going to pretend this isn't messy territory. There are legitimate security concerns about which AI systems the government uses. But those decisions need to go through proper procurement processes and technical evaluation, not political loyalty tests. The court got this one right.