Simon Willison prompts an AI: draw a pelican on a bicycle. GPT-4o spits out a flat SVG. GLM-5.1? It builds a full HTML circus—wheels spinning, beak jiggling like it’s late for a vaudeville show.
GLM-5.1. That’s the new kid from Z.ai, a 754-billion-parameter monster weighing in at 1.51 terabytes. MIT-licensed. Sitting pretty on Hugging Face, ready for your API via OpenRouter. Free as in beer—and freedom.
It’s the same size as their last GLM-5 drop. But here’s the kick: creative tasks got a brain transplant. Willison’s pelican test? Most models churn out dead images. This one goes kinetic.
“The issue is that CSS transform animations on SVG elements override the SVG transform attribute used for positioning, causing the pelican to lose its placement and fly off to the top-right. The fix is to separate positioning (SVG attribute) from animation (inner group) and use for SVG rotations since it handles coordinate systems correctly.”
That’s the model diagnosing its own mess-up. Then fixing it. Spot-on.
Why Bother with a Damn Pelican?
Pelicans aren’t just beach birds. This test probes deeper—does the AI get that code lives in a browser? SVG coords. CSS cascades. Runtime quirks. Not just pixels on a page.
Most open-weights? They barf static junk. GLM-5.1 grasps context. It’s like watching a toddler graduate to quantum physics—unexpected, a bit sloppy, but holy hell, it works.
China’s AI labs have chased benchmarks forever. Now? They’re lapping us on the fun stuff. While OpenAI polishes sales decks, Z.ai ships code that dances.
And yeah, it’s Chinese. Skeptics twitch—data poisoning? Backdoors? Please. MIT license means you audit it. Fork it. Run it local. No black-box trust falls.
Short para: Impressive.
But let’s not crown it emperor yet. First try, the pelican flew off-screen. Fixed on prompt two. Good enough for demos. Production? We’ll see.
Here’s my hot take, absent from the hype: this echoes the ’90s browser wars. Netscape open-sourced everything, forcing IE to play catch-up. GLM-5.1 could do that to closed AI giants—drop inference costs, and suddenly everyone’s baseline is a wobbling pelican, not rented pixels from Sam Altman.
Is GLM-5.1 Actually Better Than GPT-4o?
Benchmarks? It’s neck-and-neck on the charts. But pelicans reveal the gap. GPT-4o draws pretty—static, soulless. GLM animates. Understands fixes.
Size-wise, identical to GLM-5. But training tweaks flipped the script. Creative coding? Chinese edge.
Dry humor alert: Imagine Altman watching this. “Our model draws birds too!” Yeah, but does it make ‘em pedal?
Critique time. Z.ai’s PR? Subtle, no fireworks. Unlike American flash: “World’s best!” Here, they let the pelican flap. Smart.
Inference? Dropping fast. At scale, GLM-5.1 undercuts OpenAI rents. Orgs eye it for prod—license greenlights that.
One glitch: 1.51TB download. Not for your laptop. Needs serious iron. But clouds eat it.
Why Does This Matter for Open-Source Devs?
Forget cute tests. Real win: permissive license on titan params. Llama? Restricted. Mixtral? Finicky. GLM? MIT all day.
Ship it in apps. Fine-tune for niches. No lawyer calls.
Prediction: Six months, you’ll see GLM-5.1 forks in weird places—AR glasses, robot arms, indie games. Because why pay when free does better tricks?
Western open-source? We’ve idolized small models. This flips it: big, free, competent. Time to eat humble pie.
Skepticism check: Bugs lurk. Hallucinations persist. But static-to-animated jump? Tells me reasoning’s maturing.
Dense bit: Picture enterprise teams—tired of API bills, compliance nightmares. They grab GLM-5.1, quantize it to 4-bit, deploy on Kubernetes. Pelicans optional. Savings mandatory. Competitors scramble. OpenAI pivots to “premium creativity.” Too late.
And the ecosystem? Hugging Face lights up. OpenRouter queues swell. Chinese models go mainstream—ironically, powering Western startups.
Punchy: Game on.
Wander: Reminds me of Android vs iOS. Open stack wins volume. AI’s turn.
The Real Production Test
Not pelicans. It’s forklifts in warehouses, UIs in SaaS, agents debugging live. If GLM holds, it’s baseline. If not—back to rentals.
Watch costs. If 754B runs cheap, closed shops quake.
Hype callout: Blogs gush “beats GPT!” Nah. Equals on math, trumps on code context. Cherry-pick less.
🧬 Related Insights
- Read more: Transformers: The Engine Under GPT’s Hood, Minus the Hype
- Read more: MnemoPay SDK: AI Agents Get Memory and a Piggy Bank
Frequently Asked Questions
What is GLM-5.1 and where to get it?
Z.ai’s 754B open-source model, MIT-licensed on Hugging Face. API via OpenRouter.
Does GLM-5.1 really outperform GPT-4o?
On creative coding like animated SVGs, yes. Benchmarks close, but context wins.
Can I use GLM-5.1 in production?
Absolutely—MIT lets you ship, modify, anything.
How to run GLM-5.1 pelican test?
Prompt: “Draw a pelican riding a bicycle as SVG.” Watch it animate.
