Evaluating · No capacity yet

GLM-5.2: The Strongest Open-Weight Coder of 2026

Zhipu's MIT-licensed, 1M-context flagship tops open-weight coding benchmarks. Here's what's verified, where it stands, and what you can use on QCode today.

#GLM-5.2 #Open Weights #Coding #Benchmarks
🧪
Our status: evaluating — no capacity yet

GLM-5.2 is genuinely impressive, but launch-week API demand caused severe rate-limiting, and cloud access raises data-compliance considerations. We don't currently have capacity for it. If we secure reliable, compliant capacity we'll consider offering it — meanwhile we focus on models that are live today.

GLM-5.2 at a glance

MIT
Open weights

MIT-licensed on Hugging Face (zai-org/GLM-5.2) — self-hostable.

1M
1M context

A one-million-token window, four times the prior generation.

753B
753B MoE

Mixture-of-experts: 256 experts, 8 active per token (~40B active).

#1
#1 open weight

Top open-weight model on coding and agent benchmarks — still behind Claude Opus 4.8.

Verified benchmarks

Official numbers from Zhipu's release — strong, but read them in context.

62.1
SWE-bench Pro
81.0
Terminal-Bench 2.1
74.4
FrontierSWE
99.2
AIME 2026
91.2
GPQA-Diamond
76.8
MCP-Atlas

Official figures from Zhipu's release. SWE-bench Pro 62.1 edges GPT-5.5 (58.6). Note that SWE-bench Verified and Code Arena scores were not officially published — ignore fabricated numbers — and GLM-5.2 still trails Claude Opus 4.8 overall, so 'strongest open weight' is the accurate framing. Intelligence Index 51 is third-party (Artificial Analysis).

Open weights vs cloud API

Two ways to run GLM-5.2 — with very different trade-offs.

Self-host the open weights

The MIT license lets you run GLM-5.2 on your own hardware — but a 753B MoE model needs serious compute and engineering.

Vendor cloud API

Fastest to start, but launch week saw heavy rate-limiting, and routing data through a China-based provider carries compliance considerations for regulated workloads.

Where QCode stands on GLM-5.2

We've evaluated GLM-5.2 and rate it highly as an open-weight model. But supply is tight and compliance needs careful handling, so we don't offer it yet. If we can secure reliable, compliant capacity, we'll consider adding it — and we'll say so clearly here. Until then, we'd rather point you to models we can serve well today.

What you can use on QCode today

Production-ready models, available now through one key.

Claude Opus 4.8

Top-tier agentic coding.

Codex (GPT-5.5)

Terminal-native agentic workflows.

Gemini 3 Pro / 3.5 Flash

1M-context reasoning, fast multimodal.

Frequently asked questions

Is GLM-5.2 available on QCode?

Not currently. We've evaluated it and rate it highly, but supply is constrained and compliance needs care, so we don't have capacity for it yet. If that changes we'll consider adding it. Today you can use Claude Opus 4.8, Codex (GPT-5.5) and Gemini on QCode.

Is GLM-5.2 really the best coding model?

It's the strongest open-weight model — SWE-bench Pro 62.1, ahead of GPT-5.5 — but it still trails Claude Opus 4.8 among proprietary leaders. 'Best open weight' is accurate; 'best overall' is not.

Can I self-host GLM-5.2?

Yes — the weights are MIT-licensed on Hugging Face. But it's a 753B mixture-of-experts model, so self-hosting needs substantial GPU capacity and engineering.

Is GLM-5.2 free?

The open weights are free under MIT. Hosted API access is paid — third-party aggregators list around $0.95 / $3 per million input/output tokens — and was heavily rate-limited at launch.

Use a production-ready model today

Register now for Claude, Codex and Gemini — and we'll tell you here if GLM-5.2 ever joins the lineup.