A new open weights model from Chinese AI lab Z.ai dropped on June 16th under an MIT license, and the early benchmarks suggest it deserves attention. GLM-5.2 (the latest in the GLM line, a family of open weights large language models from Z.ai) is a 753 billion parameter mixture-of-experts model (a setup where only a subset of the parameters runs per token, here 40 billion of the 753 billion activate) with a 1 million token context window (the maximum amount of text the model can read in one prompt), five times larger than the previous GLM-5.1 release. It now leads the Artificial Analysis Intelligence Index v4.1 for open weights models at a score of 51, ahead of MiniMax-M3 at 44, DeepSeek V4 Pro at 44, and Kimi K2.6 at 43.
There is a trade-off. The model uses about 43,000 output tokens per Intelligence Index task, up from 26,000 for GLM-5.1. That puts it above MiniMax-M3 at 24,000, Kimi K2.6 at 35,000, and DeepSeek V4 Pro at 37,000. Token use shows up directly in API bills, so a model that is the most accurate on the leaderboard is not always the cheapest one to actually run.
What GLM-5.2 is good at
Simon Willison, who writes the long-running Weblog at simonwillison.net and has been benchmarking open weights models since the LLaMA era, ran GLM-5.2 through his usual test suite and came away impressed on the coding side. The model ranks second on the Code Arena WebDev leaderboard, behind only Claude Fable 5, on tasks that measure “front-end web development tasks, including agentic coding workflows” (a benchmark that scores models on multi-step coding problems, not just single-prompt completions).
Front-end coding benchmarks are a specific stress test because the model has to generate working HTML, CSS, and JavaScript that renders correctly in a browser. Most open weights models fall behind the closed-weight frontier here because the training data for layout and styling is harder to balance against the training data for logic and reasoning. GLM-5.2 is text-only (no image input, unlike Z.ai’s separate GLM-5V-Turbo vision family), which is the kind of constraint that usually hurts front-end benchmarks. The fact that it ranks second on this particular leaderboard is a real signal.
On the SVG generation side, the model returned a self-contained animated SVG of a pelican riding a bicycle with no broken joints or independently rotating wheels. That is a small thing, but it is the kind of detail that separates models that have actually learned animation primitives from models that have memorized training data without understanding the structure.
What GLM-5.2 is not as good at
The same benchmark run that produced the pelican produced a noticeably worse opossum. Willison’s prompt “Generate an SVG of a NORTH VIRGINIA OPOSSUM ON AN E-SCOOTER” returned a static, non-animated image, while the previous GLM-5.1 had returned an animated version that he calls “my all time favorite opossum.” The step down is sharp enough that he describes it as “such a step down from GLM-5.1.”
Token cost is the other concrete downside. At 9 different providers on OpenRouter (a routing service that lets you call many model providers through a single API), input is around $1.40 per million tokens and output is $4.40 per million tokens. That is much cheaper than GPT-5.5 ($5/$30) or Claude Opus 4.5-4.8 ($5/$25). But the model uses more tokens per task than its peers, so the per-task cost gap is smaller than the per-token price suggests. If you are running GLM-5.2 through an API for batch jobs, the actual spend will land closer to its token-usage profile than its headline price.
The vision gap also matters. If your workflow is multi-modal (you paste screenshots into the chat, ask the model to describe an image, or have it read a PDF with diagrams), GLM-5.2 will not be useful. Z.ai has GLM-5V-Turbo for that, but it is not open weights. You would have to either accept a closed-weight model for the vision portion of the pipeline or route the image work to a separate open weights model that supports vision input.
How it compares to the rest of the field
The top of the open weights leaderboard in mid-June 2026 is now GLM-5.2, with MiniMax-M3, DeepSeek V4 Pro, and Kimi K2.6 in a cluster behind it. All four are mixture-of-experts models with active parameter counts in the tens of billions, all four support long contexts, and all four are released under permissive licenses. The differences show up in benchmark scores, token efficiency, and the specific tasks each one is tuned for.
The closed-weight frontier still leads on most benchmarks. Claude Opus 4.5-4.8 and GPT-5.5 score higher than GLM-5.2 on reasoning and multi-step coding tasks, and they support vision input natively. The argument for running an open weights model is not that it is the best on benchmarks. The argument is that you can run it on your own hardware, you can audit the weights, you can fine-tune it, and you are not subject to a provider’s rate limits or pricing changes. For some teams, those properties matter more than a 5-point edge on a single benchmark.
Trade-offs
GLM-5.2 is the new leader on the Artificial Analysis Intelligence Index v4.1 for open weights models, but the model uses more tokens per task than its peers. A team that picks GLM-5.2 for accuracy pays for that accuracy in higher per-task API costs and longer time-to-first-token on responses. A team that picks MiniMax-M3 for efficiency loses some benchmark points but spends meaningfully less per task at the same provider prices.
The closed-weight frontier is still ahead on most benchmarks, and the gap on vision tasks is wider than the gap on text-only tasks. If your workflow is text-only and the model is accurate enough, GLM-5.2 is the open weights choice. If you need vision, you need a different model for that part.
The migration story is real but not free. GLM-5.2 needs around 1.5 TB of disk space for the weights, and you need a GPU with enough VRAM to host the active parameters. Self-hosting means setting up vLLM or a similar inference server, configuring API compatibility, and building the monitoring around it. The team at Z.ai released the weights under MIT, which means there is no license fee, but there is an operational cost.
What I would tell past me
If I could send a message back to the version of me that ran the first GLM-5 release, I would say three things.
- Benchmark scores are necessary, not sufficient. A model at the top of one leaderboard can fall off another. Run your own test suite on tasks that match your actual workload before you commit.
- Token cost is the second axis. Two models at the same benchmark score can have very different API bills. The headline price per million tokens hides the cost of doing the task.
- Open weights is a different argument than open source. The weights let you host, fine-tune, and audit. They do not promise reproducibility of any specific behavior, because the training data and recipes are not always public.
- Mixture-of-experts changes your hosting math. A 753 billion parameter model with 40 billion active is not the same problem as a 40 billion dense model. Memory, throughput, and latency profile all differ, and the inference server you already have for the smaller model will not just run the larger one.