Z.ai's open-source GLM-5.2 (Max) tops Code Arena's frontend leaderboard, prompting debate over its real-world performance

@ml_angelopoulos @ollama Awesome, I was pissed about the whole fable removal drama. Resubscribed to Claude for it (after 3 months of no Claude) only to have it removed in 3 days. Did Claude zero favours. Stick with ollama cloud, open code go and DeepSeek API, best and cheapest stack you can own 🚀

@ml_angelopoulos I don't trust a bench when Opus is better than gpt-5.5

@ml_angelopoulos @arena @grok who makes glm-5.2

No discussion that GLM-5.2 is very good, but beside benchmarks, people who tried it, how is it?

@0xThoughtVector @ml_angelopoulos you can use opencode yes, it has IDE extension , and an opencode desktop app (in beta) for agentic stuff BUT ZAI has its own harness called ZCode, and the main reason to use it is that they triple your plan limits if you do, which is useful because base quota aren't that generous

@ml_angelopoulos

@ml_angelopoulos show me one piece of cool software built with 5.2

@ml_angelopoulos @arena For those who are new to the race, everybody here is glazing this model, but 99% of them have never even touched it. The previous GLM models had the same Hype and yet they were shit and forgotten…

@ml_angelopoulos @ollama Is there a way to bring this GLM model into @cursor_ai? It would be awesome.

@tibo_maker @ml_angelopoulos its on the frontend

Code Arena Frontend is blind human preference on rendered UIs, so it rewards output that wins a pairwise look-test. That's a real signal for frontend specifically. It's a narrower claim than "best frontend coding model" though, since "what people vote for in a glance" and "code you'd want to maintain" aren't the same axis.

@ml_angelopoulos how do people generally use GLM models?

I don't want to use it at API pricing and I know there are plans but what app do they work in? OpenCode?

Anything good that is not a TUI?

@x509dot @ml_angelopoulos i have seen nothing but glaze posts, no demos of anything

@ml_angelopoulos @ollama With well built projects using a Good harness (Pi or Hermes) in a project folder for Front end (no one shot nonsense) I’m literally seeing practically the same results even with DeepSeek, structured small changes. Exact front end design lexicon is critical too.

@ml_angelopoulos @inductionheads This is silly because that if is doing a tremendous amount of work and this benchmark looks busted without codex.

@ml_angelopoulos It's trained by Huawei ascendant as well no Nivdia

@ollies0x @ollama It's certainly more predictable ;)

@ml_angelopoulos You shouldn't be surprised. Jensen and @nvidia warned everyone, but ain't no one listen. Useless.

@ml_angelopoulos Any ClaudeDesign alt thats using GLM? I ask Claude and it points me to Claude Design 🫩

@ml_angelopoulos Also, fable was an ensemble of models, falling back with some frequency to opus4.8 in a apecial harness that causes it to behave differently