GLM-5.3 now runs fully sealed in TeeML on 0G

Private AI, upgraded: the Private Inference tier now serves GLM-5.3. Same sealed-enclave privacy, a sharper model inside it.

On September 3, the TeeML provider on 0G Private Computer completed a scheduled upgrade from GLM 5.2 to GLM-5.3. The privacy model did not move an inch: requests still run fully sealed inside a TEE enclave, where neither 0G nor the provider can read them. The model inside the enclave got sharper. If you run Private mode pinned to glm-5.2, switch the model name to glm-5.3 and you are done.

What changed on September 3

GLM-5.3 has been live on 0G Private Computer since August 14, the day the model shipped, served through TeeTLS-attested routing from the start. What it gained this week is a second trust mode. After a ten-hour maintenance window on September 3, the Private Inference tier, the tier where the model itself runs inside a sealed TEE enclave, now serves GLM-5.3.

The changelog is short on purpose. Same Private (TeeML) privacy, a stronger model.

GLM 5.2 joined the Private Inference tier on July 7. Two months later, the same enclave runs its successor. That cadence is the point: privacy tiers only matter if they keep pace with the models people actually want to use.

What GLM-5.3 brings

GLM-5.3 is the latest flagship from Z.ai (@Zai_org), post-trained on the same 743B base as GLM 5.2 and built for coding, agentic tool use, and cyber defense. Z.ai reports 28.3 on Terminal Bench 3.0, 66.9 on DeepSWE, 28.5 on Agents' Last Exam, and 1769 on GDPVal-AA, and describes the result as matching models several times its size (Z.ai, August 2026).

One behavior worth knowing before you build on it: deep thinking is always on for this model and cannot be turned off. You control how much it thinks with reasoning_effort, which accepts low or high and, on the sealed provider, is passed inside chat_template_kwargs rather than at the top level of the request.

At launch, sealed GLM-5.3 runs $1.40 per 1M input tokens, $4.40 per 1M output tokens, and $0.35 per 1M cached input, matching the model's official list price, with reasoning tokens billed as output. Providers can adjust over time, so the model page stays the source of truth for current pricing.

What "fully sealed" actually means

TeeML is the verification mode where the model itself runs inside a Trusted Execution Environment, on Intel TDX with TEE-enabled GPUs. Your prompt enters the enclave encrypted. The response is signed inside the enclave. The host machine sees only encrypted traffic, and neither 0G nor the provider operating the hardware can read the inference data. Every enclave publishes a hardware attestation you can verify with dstack.

That is a different scope of guarantee than TeeTLS, and the difference is worth being precise about. TeeML attests the inference itself. TeeTLS attests the routing path, not the inference: a broker inside a TEE proves which upstream your prompt was sent to and how the transport was protected, while the model runs on a centralized host. TeeTLS is the smaller-scope guarantee. We covered that distinction in depth in the DeepSeek V4 Pro launch post.

On the 0G router, the two scopes map to two flags. Verified mode accepts both TeeML and TeeTLS providers and proves the response came from the real model. Private mode routes only to TeeML providers, so prompts never leave an enclave. The router privacy docs show how to enforce TeeML-only inference per API key or per request.

GLM-5.3 spent its first three weeks on 0G under TeeTLS, every routing hop attested with cryptographic proof of which upstream served the request. TeeML goes further, and 0G has run it in production before: 0GM-1.0 serves from a TEE-sealed enclave, and GLM 5.2 ran the same way from July until this upgrade.

For builders: one model name to update

Most integrations need no action at all.

If you call GLM 5.2 through the router without Private mode, nothing changes for you. GLM 5.2 keeps serving on TeeTLS providers as before.

If you use the router with Private enabled and the model pinned to glm-5.2, update the model name to glm-5.3. That is the whole migration.

If you connect to the upgraded provider directly in Advanced Mode, it now serves glm-5.3, so update the model parameter in your requests the same way.

Anthropic SDK and Claude Code users are covered as well. The sealed glm-5.3 route accepts the Anthropic Messages format, so point the Anthropic base URL at https://router-api.0g.ai, use an API key set to Private, and pin the model to glm-5.3.

To dial reasoning down for latency-sensitive calls on the sealed provider, set it inside chat_template_kwargs; a top-level reasoning_effort field is ignored there:

{
  "model": "glm-5.3",
  "chat_template_kwargs": {"reasoning_effort": "low"}
}

After the upgrade, GLM 5.2 no longer has a Private (TeeML) listing. To keep the GLM series under enclave-sealed privacy, move to GLM-5.3. If your workload needs 5.2 specifically, it stays available through TeeTLS providers, with the routing-path guarantee rather than sealed inference. Other Private/TeeML models on the platform are not affected by this upgrade.

One naming footnote so nobody migrates to the wrong SKU: glm-5.3-flash, the lightweight model listed on August 31, is a separate model. It runs on TeeTLS and is not part of this upgrade.

Frequently asked questions

What is the difference between TeeML and TeeTLS?

TeeML runs the model itself inside a TEE enclave: prompts enter encrypted, responses are signed inside, and neither 0G nor the provider can read them. TeeTLS runs a broker inside a TEE that attests the routing path to a centralized upstream, so it verifies transport, not the inference. TeeTLS is the smaller-scope guarantee.

I use GLM 5.2 without Private mode. Does anything change for me?

No. GLM 5.2 continues to serve on TeeTLS providers, and requests that never used the Private Inference tier are unaffected by this upgrade.

Can I still run GLM 5.2 on the Private Inference tier?

No. Since September 3, the Private (TeeML) listing for GLM 5.2 has been retired. For enclave-sealed privacy on the GLM series, use GLM-5.3. For GLM 5.2 specifically, TeeTLS providers remain available.

Is glm-5.3-flash part of this upgrade?

No. glm-5.3-flash is a separate lightweight model that joined 0G Private Computer on August 31 and runs on TeeTLS. This upgrade covers glm-5.3, the flagship model.

How do I verify the enclave is real?

Every TeeML enclave publishes a hardware attestation you can check with dstack, and inference on 0G is verifiable onchain. The 0G docs walk through provider verification step by step.

What does GLM-5.3 cost on the Private Inference tier?

Current pricing is listed on the model page. Prices are set per provider and can change, so the model page is the reference rather than this post.

Where can I get 0G?

get.0g.ai is the interactive guide on how and where to acquire 0G tokens, with options for developers, investors, and ecosystem participants.

Run GLM-5.3 sealed

• Try it now: pc.0g.ai/models/glm-5.3
• Enforce TeeML-only routing: router privacy docs
• The previous chapter: GLM-5.2 is live on 0G
• Follow @0G_labs for what lands next

Sources:

• Z.ai: GLM-5.3 announcement (model release, August 14, 2026)
• Z.ai: GLM-5.3 benchmarks (Terminal Bench 3.0, DeepSWE, Agents' Last Exam, GDPVal-AA)
• Z.ai tech blog: GLM-5.3 (architecture and training)
• 0G docs: router privacy (TeeML enforcement, enclave attestation)
• pc.0g.ai/models/glm-5.3 (specs and live pricing)
• GLM-5.2 is live on 0G (previous chapter, June 17, 2026)
• DeepSeek V4 Pro is live on 0G Private Computer (TeeML vs TeeTLS scope)
• 0G's first proprietary AI model (0GM-1.0, TeeML in production)