Skip to main content
Last verified: 2026-10-14
GLM-5.3 weights are released. Z.ai announced GLM-5.3 on 2026-08-14 and the weights went live on Hugging Face at zai-org/GLM-5.3 on approximately 2026-08-27. The model is also available through the Z.ai API, the GLM Coding Plan, and ZCode.
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.

GLM-5.3 release date and status

GLM-5.3 uses a custom license, not MIT. The license restricts commercial use for large companies. If your organization exceeds $10B in annual revenue over 12 months, you must complete a Z.ai security review before production deployment. GLM-4 through GLM-5.2 shipped under MIT; GLM-5.3 does not.

GLM-5.3-Flash: a separate, MIT-licensed option

Z.ai also released GLM-5.3-Flash (320B parameters), which remains MIT-licensed with no commercial restrictions. This is a separate, smaller model — not a quantization of GLM-5.3. Community GGUF uploads are available at AtomicChat/GLM-5.3-Flash-GGUF. If you need a permissively licensed GLM model at this tier, GLM-5.3-Flash is the option. If you want a smaller GLM model running today, use the GLM family page. GLM-4 9B is a one-command Ollama pull on a single consumer GPU, and GLM-4.5-Air fits a large unified-memory machine.

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 if you serve GLM locally, because the same parameter contract carries into the open-weight release and its runtime flags.

GLM-5.3 hardware requirements

Weight memory follows stored parameter count and quantization. GLM-5.3 stores 753B parameters 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.
245 GB at 2-bit fits a 256 GB Mac. An Unsloth 2-bit quantization of GLM-5.3 fits within the unified memory of a 256 GB Apple Silicon machine, leaving headroom for the OS and runtime. This is the first practical single-machine path to self-hosting GLM-5.3.
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.
At full precision (BF16), you need an ~8-GPU datacenter server. FP8 cuts that roughly in half but still requires multi-GPU infrastructure.

Fit by device

Every verdict assumes the tightest quantization that fits the target.

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. An Unsloth 2-bit build (~245 GB) does not fit a single Spark either — you would need two at 256 GB pooled, and even then headroom is tight. 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 even the 245 GB 2-bit weight floor by 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-5.3 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

GLM-5.3 weights are available on Hugging Face at zai-org/GLM-5.3. For GLM-5.3-Flash, community GGUF uploads are at AtomicChat/GLM-5.3-Flash-GGUF. There is no official Ollama tag yet. A guessed command such as ollama run glm-5.3 will not work. Note the pattern from GLM-5.2: 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.

Run it — beginner to advanced

GLM-5.2 requires --reasoning-parser glm45 and --tool-call-parser glm45 for tool calls and thinking output to separate correctly. GLM-5.3 uses the same base model, so expect equivalent flags — confirm against the model card before you standardize. See vLLM remote access.

Connect it with Tokios

An always-on local deployment is one you want to reach when you are not at the machine running it. Tokios sits after your serving runtime: it does not download the model or replace vLLM or SGLang.
1

Keep your serving runtime running

Start GLM-5.3 in vLLM or SGLang and leave it bound to loopback.
2

Pair the Tokios connector

Sign in at tokios.com/console, open the Setup tab, and click Start pairing to get a one-time claim code (TKS-XXXX-XXXX). Run the installer on the machine that runs the model:
Windows
macOS
Linux
Then approve the device on the Setup or Connectors tab.
3

Register a deployment

In the Models tab, register the upstream model id (for example zai-org/GLM-5.3) under a public deployment name like glm-5.3. Clients send the deployment name in the model field, never the local id — so swapping GLM-5.2 for GLM-5.3 later changes nothing on the client side.
4

Create an API key

In the Keys tab, click Create key. Copy the sk-tok-… key now — it is shown only once.

Call it from anywhere

GLM 5.3 Prime

GLM 5.3 Prime is a speed-optimized variant of the same 753B base model. It delivers approximately 1.5-2x faster output than standard GLM-5.3 with no change in model quality. Prime is not a separate model — it is the same weights with inference optimizations applied. Use it when you need faster throughput without changing your prompts or reasoning configuration.

Capabilities

  • Controllable reasoning. Three effort levels — low, high, max — via reasoning_effort, defaulting to max. Thinking can no longer be disabled; use low when you need faster responses.
  • Tool calling. GLM-5.3 supports structured tool calls. Serving runtimes need the correct --tool-call-parser flag (equivalent to GLM-5.2’s glm45 flag) for tool-call output to separate from thinking tokens.
  • Agentic coding. The largest gains over GLM-5.2 are on long-horizon coding tasks: Terminal Bench 3.0 jumps from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9 on Z.ai’s measurements.
  • 1M token context. 1M tokens input, ~128K output — budget KV cache memory separately from weights.
  • Vulnerability discovery. Z.ai reports the model surfaced 2,436 findings across 269 open-source projects as post-training scaled, 1,097 of them medium-to-high severity.

Caveats

  • Custom license, not MIT. Companies with >$10B revenue over 12 months need a Z.ai security review before production deployment. GLM-5.3-Flash (320B) remains MIT-licensed if you need permissive terms.
  • Benchmarks are vendor-reported. All figures above are Z.ai’s own, measured with their harness settings. Wait for independent evaluations before treating the coding scores as settled.
  • Runtime flags may change. GLM-5.2 uses --reasoning-parser glm45 and --tool-call-parser glm45; GLM-5.3 likely needs equivalent flags, but confirm against the model card and your serving runtime’s docs.
  • Thinking cannot be disabled. Any code that sets thinking.type: "disabled" must switch to enabled with reasoning_effort: "low" before upgrading to GLM-5.3.

FAQ

Yes — weights went live on Hugging Face at zai-org/GLM-5.3 on approximately 2026-08-27. The model is also available through the Z.ai API, GLM Coding Plan, and ZCode.
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.
Roughly 245 GB at Unsloth 2-bit, 377 GB at 4-bit, 755 GB at FP8, and 1.5 TB at BF16 for the 753B stored parameters, before KV cache. The 2-bit build fits a 256 GB Mac. FP8 and BF16 need multi-GPU datacenter servers. A consumer GPU is not a candidate at this tier.
Weights went live on Hugging Face at zai-org/GLM-5.3 on approximately 2026-08-27, about 2 weeks after the 2026-08-14 announcement.
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. GLM-5.3-Flash community GGUFs are available at AtomicChat/GLM-5.3-Flash-GGUF.
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.
No — GLM-5.3 ships under a custom “glm-5.3” license, not MIT. Companies with >$10B revenue over 12 months need a Z.ai security review before production use. Open weights and open source are not the same thing. GLM-5.3-Flash (320B) remains MIT-licensed if you need permissive terms.
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-5.3.

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.