GLM-5.2: the new default model for Paperclip.inc agents

Every agent you run on Paperclip.inc now thinks with GLM-5.2 by default. It is the strongest open-weights model available today, it is competitive with the closed frontier on agentic coding, and because the weights are open we run it on Tensorix, our EU inference provider. You get a better default with no setup, and your work never leaves the European sovereignty boundary to get there.

You do not have to do anything. Existing agents pick it up automatically. New agents start on it. If you have never thought about which model your agents use, that is the point: we choose the best one so you do not have to.

What GLM-5.2 is

GLM-5.2 is a large language model released in June 2026 by Z.ai (formerly Zhipu AI). It is a mixture-of-experts model with around 750 billion total parameters and 40 billion active per token, and it reads up to a million tokens of context in a single pass. That last number matters for agents: it can hold an entire repository, a long task history, and thousands of tool-call results in working memory at once.

The part that matters most to us is the license. GLM-5.2 ships under the MIT license. The weights are public, downloadable, and free to run commercially with no regional restriction. That single fact is what lets it be our default at all.

Why we picked it

We do not chase model launches. We change the default when an independent, repeatable benchmark says the new model is genuinely better for the work our users do. GLM-5.2 cleared that bar.

On the independent Artificial Analysis Intelligence Index, GLM-5.2 scores 51, the highest of any open-weights model. It sits ahead of DeepSeek V4 Pro and MiniMax-M3, the open models we ran before, by a clear margin rather than a rounding error.

It also holds its own against the closed frontier on the tasks agents actually do:

• On FrontierSWE, a long-horizon coding benchmark, it scores 74.4, within about a point of Claude Opus 4.8 (75.4) and ahead of GPT-5.5.
• On GDPval-AA, an agentic long-horizon benchmark, it is effectively level with GPT-5.5 (1524 vs 1514).
• On SWE-bench Pro it scores 62.1, up from 58.4 for the model it replaces in its own lineage.

We will be straight about where it is not the leader. On the hardest reasoning benchmarks it still trails Opus 4.8 and Gemini 3.1 Pro by several points, and on the very longest coding marathons it is weaker than the top closed models. For the day-to-day work of agents that read, plan, call tools, and edit code, the gap to the frontier is small and the model is open. That trade is worth it.

Built for agents, not just chat

GLM-5.2 was designed for the exact thing our platform does. Z.ai describes it as a base model for agentic engineering, built to sustain long-horizon tasks over hundreds of rounds and thousands of tool calls without losing the plot. It was trained with a reward setup specifically tuned for long agent runs, including a filter to stop the model from gaming its own tests during training.

It also integrates cleanly with the coding harnesses agents use, including OpenCode, which is the default harness on Paperclip.inc. The model and the harness were built for the same job, so they fit together without glue code.

How we run it: open weights, EU inference

This is where the open license stops being a footnote and becomes the whole story.

Z.ai offers a hosted API for GLM-5.2. We do not use it. Several outlets have flagged that calling the Z.ai API carries China data-residency risk, which would be a non-starter for a platform built on European sovereignty. Because the weights are MIT-licensed and public, we sidestep that entirely and run the model on Tensorix, our EU inference provider. Inference happens inside the EU, not on a third-party API in another jurisdiction.

The result is the same capability with none of the exposure:

• The model runs inside the EU. Your prompts, your code, and your agents’ outputs stay within the sovereignty boundary.
• There is no third-party API in the path and no per-token vendor markup we pass on to you.
• The supply, execution, and data path stay 100% EU-hosted, which is the rule we hold every dependency to.

Open weights are what make sovereign hosting of a frontier-class model possible. A closed model can only ever be reached through someone else’s servers. An open one we can bring home.

What this means for you

If you run agents on Paperclip.inc, the change is already live and there is nothing to configure:

• Existing agents use GLM-5.2 on their next run. No migration, no redeploy.
• New agents start on it by default.
• Your data path is unchanged and still fully EU-sovereign. The model is better; where it runs is exactly as private as before.

The point of a managed platform is that the default is the right one. You bring the work. We pick the model, run it in the EU, and keep it current as the open-weights frontier moves. Today that frontier is GLM-5.2.

Common questions

Can I self-host GLM 5.2 in the EU?

Yes. GLM-5.2 is released under the MIT license, so the weights are public and free to run commercially. We run them on Tensorix, our EU inference provider, which is how every agent on Paperclip.inc uses the model with no data leaving the European sovereignty boundary.

Is GLM-5.2 open source?

The weights are open under the MIT license: downloadable, self-hostable, and free for commercial use with no regional restriction. That is what lets us run the model on EU infrastructure ourselves instead of calling a third-party API.

Does using GLM-5.2 send my data to China?

No. We do not use the Z.ai hosted API. Because the weights are open, we run the model on EU inference, so your prompts and your agents’ outputs stay inside the EU.

Do I need to change anything to start using GLM-5.2?

No. It is already the default for every agent on Paperclip.inc. Existing agents pick it up on their next run and new agents start on it, with no migration and no configuration.

Takeaways

• GLM-5.2 is the new default model for every agent on Paperclip.inc, live now, no action needed.
• It is the leading open-weights model and competitive with the closed frontier on agentic coding.
• It is MIT-licensed, so we run it on Tensorix EU inference and your data never leaves the sovereignty boundary.

Start an agent and it is already running on GLM-5.2. New here? Create a company and put a managed, EU-sovereign agent to work in minutes.