Skip to main content
Last verified: 2026-08-15
GLM-5.3 is released, but not yet as weights. Z.ai announced it on 2026-08-14 and it is live through the Z.ai API, the GLM Coding Plan, and ZCode. The checkpoint is not public: Z.ai says weights ship about two weeks after launch, once safety evaluation and hardening are complete. This page covers what that means for a local deployment, and what you can size and prepare now.
The unusual part of this release is how much you already know about running it. GLM-5.3 uses the same base model as GLM-5.2 — every gain comes from post-training — so the parameter count, memory profile, and serving path carry over unchanged. You can size hardware today with confidence rather than waiting for the model card.

​
GLM-5.3 release date and status

If you want a GLM model running today, use the GLM family page. GLM-4 9B is a one-command Ollama pull on a single consumer GPU, GLM-4.5-Air fits a large unified-memory machine, and GLM-5.2 shares this model’s base if you want the closest thing to GLM-5.3 you can self-host now.

​
What actually changed from GLM-5.2

Z.ai describes GLM-5.3 as a post-training release: the pretrained base is identical, and the improvements come from scaled reinforcement learning on long-horizon task environments. The reported gains are large, particularly on agentic coding. Z.ai also reports better token efficiency, not just higher scores: on its in-house Z.ai Code Bench at max effort, GLM-5.3 reaches 34.5% using roughly 75K output tokens per task, against 23.4% at 96K for GLM-5.2. All figures above are the vendor’s own, measured with their harness settings — treat them as a direction of travel until independent evaluations land. The release also reports a sharp jump in vulnerability-discovery capability, which Z.ai says grew faster than expected as post-training scaled. Working with security teams, the model surfaced 2,436 findings across 269 open-source projects, 1,097 of them medium-to-high severity, tracked in a public disclosure ledger. That matters here only as a capability signal; it does not change how you serve the model.

​
The API change that breaks upgrades

Thinking can no longer be disabled. GLM-5.2 accepted thinking.type: "disabled"; GLM-5.3 rejects it, and the request fails. If your application runs GLM with thinking off, you must switch to thinking.type: "enabled" with reasoning_effort: "low" before changing the model id.
GLM-5.3 exposes three reasoning levels through reasoning_effort — low, high, and max — defaulting to max, which Z.ai recommends for coding work:
This is worth planning for even if you only use GLM locally later, because the same parameter contract usually carries into the open-weight release and its runtime flags.

​
GLM-5.3 hardware requirements

Weight memory follows stored parameter count and quantization. Because the base model is unchanged, GLM-5.3 stores the same 753B parameters as GLM-5.2 in a mixture-of-experts layout: Those are weights only. Read them as the floor beneath which nothing runs, not as a requirement — KV cache, runtime buffers, and system reservation all come on top.
In a mixture-of-experts model every expert weight stays resident even though only a fraction activates per token. Memory scales with total parameters, not active ones, so do not size a machine from an active-parameter count. See how inference works.

​
Fit by device

Every verdict assumes a 4-bit build at GLM-5.2’s stored size.

​
Can a DGX Spark run GLM-5.3?

Not one, and probably not two. A DGX Spark holds 128 GB of unified memory against a ~377 GB weight floor at 4 bits. Three Sparks clear that floor on paper at 384 GB, but leave about 7 GB total for the operating system, runtime buffers, and KV cache across the whole cluster — unusable in practice. Four Sparks at 512 GB is the first configuration with genuine headroom. See DGX Spark model compatibility for the sizing method.

​
Can a 24 GB GPU run GLM-5.3?

No, and offloading will not rescue it. A 24 GB card is short of a 377 GB weight floor by more than an order of magnitude. Streaming weights from system RAM or disk at that ratio collapses throughput to the point of uselessness — see memory and offloading for why. If you have a single consumer GPU and want a GLM model, run GLM-4 9B at 5.5 GB:
Then reach it from anywhere with Tokios. For a fuller picture of what your hardware can hold, use what size model fits your GPU?.

​
Context is a separate budget

The figures above cover weights only. GLM-5.2 is natively 1M tokens, and a context that long can cost more memory than a mid-size model’s entire weight set. Budget for it separately — see KV cache and context.

​
GLM-5.3 download and Ollama

There is no GLM-5.3 download or Ollama tag yet. A guessed command such as ollama run glm-5.3 will not work, and community uploads appearing before Z.ai’s files are not official releases. Note the pattern from GLM-5.2, which likely carries over: it is listed on Ollama only as glm-5.2:cloud, which routes to Z.ai’s hosted infrastructure rather than downloading weights. At this size class, expect a cloud tag rather than a local pull.

​
Prepare a local deployment

Because the base model is unchanged, you can prepare almost everything before the weights land:
  1. Decide honestly whether you are self-hosting this tier or using a hosted endpoint. At a ~377 GB weight floor, most people are using a hosted endpoint — and Tokios fronts either.
  2. If you are self-hosting, plan multi-GPU serving with vLLM or SGLang. GLM-5.2 needs --reasoning-parser glm45 and --tool-call-parser glm45; expect GLM-5.3 to need equivalent flags before tool calls and thinking output separate correctly.
  3. Audit any code that sets thinking.type: "disabled" now — that path stops working on GLM-5.3 regardless of how you serve it.
  4. Pair the Tokios connector with whatever OpenAI-compatible endpoint you land on.
  5. Register a deployment name in the dashboard. Clients send that deployment name in the model field, not the upstream model id — so when you swap GLM-5.2 for GLM-5.3 later, nothing on the client side changes.
That last point is the reason to set this up before release rather than after: the swap becomes a one-line change on the serving machine, and every agent and teammate pointed at your endpoint keeps working.

​
What is still pending

Z.ai has published the announcement, benchmarks, and API contract. Still outstanding:
  • the weights, model card, and license text;
  • the exact repository, upstream model id, and file sizes;
  • verified vLLM, SGLang, and Ollama commands;
  • tested hardware guidance and Tokios connector examples.

​
FAQ

Yes, as of 2026-08-14 — but through the API, GLM Coding Plan, and ZCode rather than as downloadable weights. Z.ai says the checkpoint follows roughly two weeks after launch, once safety evaluation and hardening finish. Until then the newest GLM you can self-host is GLM-5.2, which shares GLM-5.3’s base model.
Post-training only — the pretrained base is identical, so the size and hardware profile are unchanged. The gains are concentrated in agentic coding and long-horizon tasks: Terminal Bench 3.0 goes from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9 on Z.ai’s own measurements. There is also a breaking API change — thinking can no longer be disabled.
The same as GLM-5.2, because it is the same base model: roughly 377 GB at 4-bit, 753 GB at 8-bit, and 1.5 TB at BF16 for the 753B stored parameters, before KV cache. That is multi-GPU server or multi-node territory, not a single workstation and not one DGX Spark. A consumer GPU is not a candidate at this tier.
Z.ai’s announcement says approximately two weeks after the 2026-08-14 launch, pending safety evaluation and hardening — so late August 2026, with the Hugging Face listing already marked “coming soon”. No exact date has been committed, so treat it as the vendor’s stated plan.
Not today — no tag exists. Expect a cloud tag rather than a local pull when it ships: GLM-5.2 appears on Ollama only as glm-5.2:cloud, which routes to Z.ai’s hosted infrastructure instead of downloading weights. For a GLM model you can actually pull now, ollama run glm4 gets you the 9B at 5.5 GB.
Only if you are not disabling thinking. GLM-5.3 rejects thinking.type: "disabled" and the request fails, so any client relying on non-thinking mode has to move to enabled with reasoning_effort: "low" first. Reasoning depth is otherwise controlled by low, high, or max, defaulting to max.
Z.ai has committed to releasing the weights, and GLM-4 through GLM-5.2 shipped under MIT, but GLM-5.3’s license has not been published yet. Open weights and open source are not the same thing — read the license when it lands, particularly for commercial use.
Serve it on any OpenAI-compatible runtime, then pair the Tokios connector so the machine holding the GPUs needs no inbound ports. You call https://api.tokios.com with an sk-tok-… key from any OpenAI- or Anthropic-compatible client, sending your deployment name in the model field rather than the upstream model id.

GLM family

Run a verified GLM model today, from the 9B up to GLM-5.2.

Model fit

Estimate the model and context size your hardware can hold.

Quickstart

Pair the connector and register a local deployment.

Quantization and hardware

What a model’s release formats mean for your hardware.