Model bench · Zhipu AI / Z.ai
Deploy GLM-5.3 on-premise.
GLM-5.3 is the strongest open-weight agentic coding model as of September 2026: a 753-billion-parameter mixture-of-experts on the same base as GLM-5.2, with a 1M-token context window and post-training that lifted Terminal-Bench 3.0 from 4.6 to 28.3 and Terminal-Bench 2.1 to 88.2. Weights shipped on Hugging Face on August 28, 2026. It is an in-place upgrade for every GLM-5.2 rack we run — with one licensing change procurement should read before signing.
Specifications
GLM-5.3 at a glance.
| Spec | GLM-5.3 |
|---|---|
| Architecture | 753B-parameter MoE · 78 layers · 256 routed experts, 8 active + 1 shared per token (~40B active, third-party estimate) |
| Attention | DeepSeek-style sparse attention (DSA) with indexer — same weight shape as GLM-5.2 |
| Context window | 1,048,576 tokens (config); 128K max output reported |
| Modalities | Text-only — pair with GLM-5.3-Flash or Qwen3.8 for vision |
| License | GLM-5.3 License — MIT-style grant plus a security-review clause for model-as-a-service operators above US$10B group revenue |
| Released | Announced Aug 14, 2026 (Coding Plan) · open weights Aug 28, 2026 · zai-org/GLM-5.3 |
| Benchmarks (model card) | Terminal-Bench 2.1 88.2 · Terminal-Bench 3.0 28.3 · DeepSWE v1.1 66.9 · CyberGym 84.5 · +50% over GLM-5.2 on Z.ai Code Bench |
| Serving | vLLM · SGLang · Transformers · TokenSpeed · KTransformers · Unsloth · Ascend NPU |
Z.ai’s own summary of the release: “Scaling post-training is all we did for GLM-5.3.” The base model, tokenizer, and tensor layout match GLM-5.2, which is why an existing GLM-5.2 deployment upgrades by swapping checkpoints — no re-architecture, no new serving stack.
Hardware requirements
Sizing the deployment honestly.
Quantized single-node tier
At 2–4 bits the full 753B checkpoint lands between roughly 245 GB and 400 GB of accelerator memory, which fits a single 8-GPU node with 96–192 GB parts (third-party sizing; we validate on your workload). This is the practical tier for a firm-wide coding and agent workhorse when you want the full model rather than Flash.
Full-precision rack tier
BF16 weights are about 1.5 TB and FP8 about 810 GB before KV cache, so full-precision serving with headroom for long contexts is a two-node or dense-rack build. Per-token cost falls as utilization rises, and overnight agentic batch runs are free at the margin.
Capabilities
What it’s best at.
GLM-5.3 is a coding and autonomous-agent specialist. Its post-training targeted exactly the workloads regulated engineering teams want to keep inside the perimeter: repository-scale software engineering, terminal-driven task completion, structured tool use over long horizons, and security research — Z.ai reports the model found 2,436 vulnerabilities across 269 projects during testing, and it more than doubles GLM-5.2 on exploitation benchmarks.
For most organizations the decision is GLM-5.3 versus GLM-5.3-Flash: Flash is MIT-licensed, multimodal, less than half the memory, and within a few points on the agentic benchmarks. We deploy the full model where the last few points on coding autonomy justify the rack, and Flash everywhere else.
Deployment patterns
How we deploy it.
In-place upgrade from GLM-5.2. Same vLLM or SGLang stack, same tensor-parallel layout, new checkpoint. We pin and checksum the weights, re-run your evaluation harness, and cut over behind the existing gateway — typically a change window, not a project.
The engineering-org coding node. One 8-GPU node serving quantized GLM-5.3 to internal coding agents and CI, with repository access scoped by role and every tool call logged. Air-gapped operation is fully supported; nothing about the model requires a network.
License review. The GLM-5.3 License grants use, copy, modify, distribute, sublicense, sell, run, deploy and fine-tune. The one added clause applies to model-as-a-service operators whose group revenue exceeds US$10B over any consecutive twelve months, who must pass a Z.ai security review before commercial use. Embedding the model in your own product or running it internally does not trigger it. We include the license analysis in every deployment file.
Questions we get
Frequently asked questions
How many GPUs does GLM-5.3 need?
The checkpoint is 753B parameters — about 1.5 TB at BF16, roughly 810 GB at FP8, and between about 245 GB and 400 GB at 2–4-bit quantization (third-party sizing). A quantized deployment fits a single 8-GPU node with 96–192 GB accelerators; full-precision serving with long-context headroom is a two-node or dense-rack build. Only ~40B parameters activate per token, so compute per token is modest and a well-fabriced node sustains real concurrency. The sovereignty assessment converts your workload into an exact bill of materials.
Is GLM-5.3 still MIT-licensed like GLM-5.2?
No. GLM-5.3 ships under the “GLM-5.3 License”: the same MIT-style grant (use, copy, modify, distribute, sublicense, sell, run, deploy, fine-tune) plus one clause requiring model-as-a-service operators with more than US$10B in group revenue over any consecutive twelve months to pass a Z.ai security review before commercial use. Internal deployment and embedding the model in your own product do not trigger it. GLM-5.3-Flash, released the same week, remains plain MIT.
Should we upgrade from GLM-5.2 to GLM-5.3?
For coding and agentic workloads, yes — the gains are large (Terminal-Bench 3.0 from 4.6 to 28.3; +50% on Z.ai’s code bench) and the upgrade is a checkpoint swap on the same serving stack. For general assistant, document, and vision work, GLM-5.3-Flash is the better move: MIT license, multimodal, 18B active parameters, and under half the memory. Most of our GLM-5.2 racks are moving to Flash, with the full GLM-5.3 reserved for engineering organizations.
How does GLM-5.3 compare with Kimi K3 and DeepSeek V4?
On agentic coding they are close: Kimi K3 posts Terminal-Bench 2.1 88.3 and DeepSWE 67.5, DeepSeek V4-Pro-0813 posts 87.9, and GLM-5.3 posts 88.2 and 66.9. The differences are elsewhere — K3 is 2.8T parameters with native vision and needs supernode-class hardware; DeepSeek V4-Pro is MIT-licensed at 1.6–1.7T parameters; GLM-5.3 is the smallest of the three at 753B and upgrades an existing GLM-5.2 rack in place.
Upgrade the workhorse without re-racking. Size it.
The two-week sovereignty assessment sizes the hardware against your real workloads and hands you a written architecture with a cost model — before you buy a single GPU.
Book a sovereignty assessment