How many GPUs do you need to run GLM-5.3?

For quantized production serving, one 8-GPU node with 96–192 GB accelerators: the 753B checkpoint is roughly 245 GB at 2-bit and about 400 GB at 4-bit (third-party sizing), plus KV-cache headroom. FP8 (~810 GB) and BF16 (~1.5 TB) serving with long-context headroom is a two-node or dense-rack build. Only ~40B parameters activate per token, so per-token compute is modest and one well-fabriced node sustains real concurrency.

Is GLM-5.3 MIT-licensed like GLM-5.2?

No. GLM-5.3 ships under the GLM-5.3 License: an MIT-style grant (use, copy, modify, distribute, sublicense, sell, run, deploy, fine-tune) plus one added 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, is plain MIT.

Can I upgrade an existing GLM-5.2 deployment to GLM-5.3?

Yes, and that is the point of the release. Z.ai states the base model is unchanged and all gains come from scaled post-training, so the tensor layout, tokenizer and serving configuration carry over. The upgrade is a checkpoint swap on the same vLLM or SGLang stack, followed by a re-run of your evaluation harness.

GLM-5.3 or GLM-5.3-Flash — which should we deploy?

GLM-5.3 when the workload is agentic coding, terminal-driven task completion, or security research and the extra points justify an 8-GPU node. GLM-5.3-Flash for everything else: it is MIT-licensed, natively multimodal, 320B/18B active, and its 4-bit build fits in about 200 GB. Most GLM-5.2 racks are re-provisioning to Flash; engineering organizations are taking the full GLM-5.3.

GLM-5.3 GPU Requirements: Sizing the 753B Coding Flagship

GLM-5.3 is the strongest open-weight coding and agent model you can download as of September 2026, and its GPU requirements follow from two numbers: 753 billion total parameters and roughly 40 billion active per token. Z.ai (Zhipu AI) announced the model on August 14, 2026 inside its GLM Coding Plan and published the weights on Hugging Face as zai-org/GLM-5.3 on August 28, 2026. Memory must hold the whole mixture-of-experts — about 1.5 TB at BF16, roughly 810 GB at FP8, and between 245 GB and 400 GB at 2–4-bit quantization (third-party sizing) — while per-token compute stays in mid-size-model territory. In practice that means one 8-GPU node for quantized production serving and a two-node or dense-rack build for full precision. Because the base model is identical to GLM-5.2, an existing GLM-5.2 rack upgrades by swapping checkpoints. Here is the sizing, the benchmark picture, and the one licensing change procurement needs to read.

What is GLM-5.3, exactly?

Attribute	GLM-5.3
Developer	Zhipu AI (Z.ai)
Architecture	Mixture-of-experts: 753B total; 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 (`max_position_embeddings`); 128K max output reported
Modalities	Text-only
License	GLM-5.3 License (MIT-style grant + one review clause for very large MaaS operators)
Released	Announced Aug 14, 2026 · weights Aug 28, 2026
Hugging Face	`zai-org/GLM-5.3`
Serving frameworks	vLLM, SGLang, Transformers, TokenSpeed, KTransformers, Unsloth; Ascend NPU supported
API price (hosted, for comparison)	$1.40 / 1M input · $0.26 cached · $4.40 / 1M output — identical to GLM-5.2

Z.ai's own description of the release is unusually plain: "Scaling post-training is all we did for GLM-5.3." The pre-trained base is GLM-5.2's. Everything that changed — the benchmark jumps, the agentic behaviour, the security-research capability — came from reinforcement learning and supervised post-training on the same weights shape. That is why the model card's config.json reads like GLM-5.2's, and why the practical upgrade path is a file swap rather than a re-architecture.

How good is it, and at what?

The model card's numbers are specific and, for a coding shop, decisive:

Benchmark	GLM-5.3	GLM-5.2	Notes
Terminal-Bench 2.1	88.2	—	Kimi K3: 88.3 · DeepSeek V4-Pro-0813: 87.9
Terminal-Bench 3.0	28.3	4.6	The largest single jump in the release
DeepSWE v1.1	66.9	—	Kimi K3: 67.5
CyberGym	84.5	—	Security research
ExploitBench	54.4	—	"More than doubles GLM-5.2"
Z.ai Code Bench (in-house)	+50% vs GLM-5.2	baseline

Z.ai also reports the model found 2,436 vulnerabilities across 269 projects during testing. For a regulated engineering organization that is the headline capability and the headline governance question in one: a model this good at finding exploitable code is exactly the kind of tool that should run inside the perimeter, with every tool call logged, rather than through a hosted API.

What GLM-5.3 is not: multimodal. It is text-only, like GLM-5.2. Vision workloads pair it with GLM-5.3-Flash or the Qwen3.8 family behind the same gateway.

How much GPU memory does each precision tier need?

All 753B parameters must sit in accelerator memory even though only ~40B activate per token — experts wait in VRAM to be routed to. Weight footprints follow from the parameter count; add 15–30% for KV cache, activations and serving overhead, more if you serve 1M-token contexts concurrently.

Precision	Weights (approx.)	Aggregate memory target	Typical configuration	Use case
BF16	~1.5 TB	1.8 TB+	16+ GPUs, two nodes or dense rack	Maximum quality; evaluation baselines
FP8	~810 GB	1.0 TB+	8 × 141 GB or two 8-GPU nodes with 96 GB parts	Production quality with long-context headroom
4-bit	~400 GB	480–520 GB	8 × 80/96 GB or 4 × 141 GB	Production serving for an engineering org
2-bit	~245 GB	300 GB+	4 × 80 GB or 2 × 141 GB	Evaluation; latency-tolerant batch

The 2-bit and 4-bit figures are third-party estimates rather than Z.ai's own, and low-bit quantization of a model whose value is precise code generation needs to be validated on your repositories before you trust it in production. Our default recommendation for a firm-wide coding node is 4-bit or FP8 on eight accelerators; we treat 2-bit as an evaluation tier.

Which node layout works?

Eight-GPU node, 4-bit (~400 GB). The practical production tier. Tensor-parallel across eight 80–96 GB accelerators leaves roughly 240–370 GB for KV cache — enough for real concurrency at 128K–256K contexts and single-session million-token runs.
Eight-GPU node, FP8 (~810 GB) on 141 GB parts. Near-full quality with headroom, if you own or are buying current-generation accelerators.
Two nodes, BF16 (~1.5 TB). For evaluation baselines and for organizations that refuse to quantize the model whose output is code. Pipeline- or expert-parallel across nodes; inter-node fabric matters.
Ascend NPU. Z.ai lists Ascend deployment support on the card, which matters for organizations whose procurement is constrained away from a single accelerator vendor.

Throughput follows active parameters, not total: ~40B active parameters is mid-size-model compute, so a saturated 8-GPU node serves an engineering organization's coding agents and CI at a per-token cost that falls as utilization rises. Overnight agentic batch — repository migrations, test generation, security scans — runs at zero marginal cost.

Upgrading from GLM-5.2: what actually changes

Nothing about the serving stack. The same vLLM or SGLang configuration, the same tensor-parallel layout, and the same tokenizer carry over; you download the new checkpoint, verify it against the published hashes, re-run your evaluation harness, and cut over behind the gateway. We schedule it as a change window, not a project. The one thing to re-check is behaviour on your own evaluation set: post-training this aggressive changes how a model follows instructions and uses tools, and a coding agent's prompts tuned against GLM-5.2 may need adjustment.

The license clause procurement will ask about

GLM-5.2 was MIT. GLM-5.3 is not. The GLM-5.3 License keeps the MIT-style grant — use, copy, modify, distribute, sublicense, sell, run, deploy, fine-tune — and adds one clause: model-as-a-service operators whose group revenue exceeds US$10 billion over any consecutive twelve months must pass a Z.ai security review before commercial use (contact glmlicense@z.ai). Embedding the model in an end-user product does not trigger the review; relaying requests to a third-party host does not either. For essentially every enterprise and sovereign deployment the clause never applies. It is still a change from "MIT, no questions," and a procurement team reading the file for the first time will notice it. Two practical consequences:

Put the license analysis in the deployment file alongside the checksums and the evaluation results. Regulators and auditors increasingly ask for it.
Where the license must be unambiguous — a public-sector buyer, a firm with a strict open-source policy — GLM-5.3-Flash is plain MIT and, for most non-coding workloads, the better model anyway.

GLM-5.3 versus GLM-5.3-Flash

	GLM-5.3	GLM-5.3-Flash
Base	GLM-5.2 base, scaled post-training	New hybrid sparse + linear attention base
Parameters	753B / ~40B active	320B / 18B active
Modalities	Text-only	Text, image, video, visual documents
Context	1M	1M (128K output)
License	GLM-5.3 License	MIT
4-bit footprint	~400 GB	~200 GB
Terminal-Bench 2.1	88.2	84.3
DeepSWE v1.1	66.9 (card)	63.4 (card)
Hosted price	$1.40 / $4.40 per 1M	$0.15 / $0.50 per 1M

Head-to-head DeepSWE figures vary across sources and should be treated as indicative. The decision rule we use: GLM-5.3 when the workload is agentic coding or security research and the last few points justify an 8-GPU node; Flash for the general assistant, document, vision and agent-runtime workloads that make up most of an organization's volume.

Where it sits on the bench

The end of August 2026 produced five frontier-class open-weight releases in nine days, and GLM-5.3 is the coding specialist among them. On agentic coding it is within a point of Kimi K3 (2.8T parameters, supernode-class hardware, Kimi K3 License) and DeepSeek V4-Pro-0813 (1.6–1.7T parameters, MIT). It is the smallest of the three and the only one that upgrades an existing rack in place. The full comparison is in our open-weight enterprise bench, and the August 2026 release roundup covers the rest of the field.

How we deploy it

The on-premise LLM deployment engagement takes GLM-5.3 from the assessment to a hardened node inside your perimeter: pinned and checksummed weights, an evaluation harness run against your repositories, SSO and role-scoped repository access for the agents, and full audit logging of every tool call. For engineering and manufacturing organizations, where the code base and the CAD-adjacent archive are the trade secrets, air-gapped operation is the default rather than the exception — and nothing about GLM-5.3 requires a network. The GLM-5.3 model page has the specification table; the GLM-5.3-Flash hardware guide covers the sibling most organizations will run beside it.

The two-week sovereignty assessment converts your repositories, concurrency and context profile into an exact bill of materials — and tells you honestly whether you need the full GLM-5.3 or whether Flash on a smaller node clears your bar.