GLM-5 Drops: 744B Open-Weight Model Under MIT License Shakes Up the Coding Leaderboard

Good morning and welcome to your AI News Daily Briefing for February twelfth, twenty twenty-six. I'm Alex, joined as always by Sam — and we have got a packed show today.

Yeah, big day. We've got a massive open-weight model drop, a talent exodus at xAI, critical security patches you need to know about, and honestly the whole vibe today is open-source models just flexing on the closed labs.

So let's get right into it. The hero story today: GLM-5, codenamed Pony Alpha, just dropped. It's a seven hundred forty-four billion parameter model, fully open-weight, under the MIT license. And developers are losing their minds over it.

Okay, so the parameter count is wild — seven forty-four B is enormous, you're not running this on your MacBook — but the thing that actually matters here is MIT license. That's the most permissive license you can get. No commercial restrictions, no fine-print, no calling your lawyer. You just use it.

Exactly. And the early reactions from people who've actually used it for code generation are really strong. Theo is ranking it alongside Opus four point six and Codex five point three, claiming it outperforms Gemini. Benchmarks are still thin, but the practitioner enthusiasm is real.

That's interesting because if those claims hold up, this fundamentally changes the build-versus-buy decision for any company doing code-heavy AI workflows. Why pay per-token to an API provider when you can run a frontier-class model yourself with zero licensing overhead?

Right, and what's wild is the timing. This drops the same week that xAI is falling apart, Apple delays Siri again, and Anthropic is busy trying to grow its free tier. The closed labs are playing defense while open-source just ships.

It's a pattern we keep seeing accelerate. And it ties directly into the Modal Labs fundraise we'll talk about — because the infrastructure to actually self-host these models is becoming a huge business in itself.

Speaking of things falling apart — let's talk about xAI. Nine engineers gone in a single week, including two co-founders. This is happening amid delays to Grok four point two, which was originally supposed to ship late twenty twenty-four.

Nine in a week is not turnover, that's a crisis. Two co-founders leaving — that's the kind of thing that makes you wonder if the roadmap is even executable anymore. Especially when every competitor is shipping aggressively.

Meanwhile Anthropic is making some interesting moves. Claude's free tier just got a serious upgrade — file creation, third-party connectors, skills, and automatic context compaction. That compaction feature is huge — it summarizes earlier context so you can have much longer conversations without hitting the wall.

Smart play. Get people hooked on real workflow features for free, then convert them. Though I'd note Anthropic also lost another safety researcher this week who left warning about catastrophic risks. That's becoming a pattern across all the frontier labs.

It is. And then there's Apple — Gemini-powered Siri features pushed back yet again, now beyond iOS twenty-six point four. At this point, every month of delay is a month where competitors are embedding AI assistants deeper into user habits.

I honestly think this is an architecture problem, not a timeline problem. If it were easy to bolt a frontier model onto Siri, they would have done it by now.

One more AI story worth flagging — OpenClaw just hit a hundred eighty thousand GitHub stars. Lex Fridman interviewed the creator, and developers are building multi-agent systems on consumer hardware. One user reportedly ran a six-agent team on a Mac Mini.

A hundred eighty K stars is insane. That's hitting the sweet spot where the framework is powerful enough to be useful but accessible enough that individual developers can actually run it. That's the magic combination.

Let's shift to dev tools. A couple of things caught my eye. Devin code review just crossed forty thousand daily uses and added one-click merge fixes. And Windsurf launched an arena-mode leaderboard that pulled forty thousand votes in its first week.

The Windsurf arena thing is really interesting to me. We've needed a Chatbot Arena equivalent for code generation for a while now. Synthetic benchmarks only tell you so much — having real developers vote on real outputs is way more signal.

Also, GitHub went down yesterday, and people noticed the status page conveniently doesn't show historical uptime data. A third-party tracker had to fill the gap. Not a great look for the world's most critical code hosting platform.

Yeah, that's embarrassing. If you're the backbone of the developer ecosystem, transparency on uptime shouldn't be optional. Developers notice these things.

On the funding side — Modal Labs is raising at a two and a half billion dollar valuation for AI inference infrastructure. General Catalyst leading. And this ties right back to our hero story.

Totally. If open models like GLM-5 make self-hosted inference more attractive, then the companies making that inference easy to run become the picks-and-shovels play. Two and a half billion makes a lot of sense in that context.

Okay, rapid-fire security segment because this is urgent. Microsoft is warning about active zero-day exploits targeting Windows and Office. These allow complete system takeover via malicious links, and they are being exploited right now in the wild. Patch today.

And if you're on Apple devices, iOS twenty-six point three patches over thirty-five vulnerabilities. Plus there's a wild one — Microsoft patched a flaw in Notepad where Markdown rendering could enable remote code execution. Even the simplest tools can become attack vectors when you add features.

That Notepad one is almost poetic. All right, quick hits. AMD is partnering with LiquidAI for on-device AI models. Perplexity shipped a memory feature that works across multiple reasoning models. And Google managed to both release and then pull Android seventeen Beta one on the same day.

The Android beta thing is just baffling. Also worth noting — Apple is adding a transfer-to-Android tool in iOS twenty-six point three, which is clearly an EU regulatory move, but still surprising to see from Apple.

So to wrap it up, Sam — the dominant pattern today is unmistakable. The center of gravity in AI is shifting toward open models and self-hosted infrastructure. GLM-5 under MIT at seven forty-four B, Modal raising at two and a half billion, OpenClaw at a hundred eighty K stars — it all points the same direction.

And on the other side, the closed labs are fighting fires — xAI losing co-founders, safety researchers leaving Anthropic, Apple unable to ship Siri. The builders betting on open, composable infrastructure are the ones setting the pace right now.

That's your briefing for February twelfth. Links to everything we mentioned are in the show notes. Patch your systems, check out GLM-5 if you're building code tooling, and we'll see you tomorrow.

See you then. Happy building, everyone.