TL;DR: GLM-5.3 is Z.ai’s post-training refresh of GLM-5.2 — same 744B/40B MoE base, same 1M context, noticeably better agentic coding scores (Terminal-Bench 2.1: 88.2 vs 81.0). The catch is the license: GLM-5.2 was MIT, GLM-5.3 ships under a bespoke “GLM-5.3 License” that adds a security-review requirement for Model-as-a-Service providers above $10B revenue. For almost every self-hoster that clause will never trigger — but it is no longer MIT, and your legal review should know that. Hardware requirements are unchanged and still brutal: this is an API-or-rented-pod model for nearly everyone.

	GLM-5.3	GLM-5.2	GLM-5.3-Flash
Parameters	744B total / ~40B active (MoE)	744B / ~40B active	320B total / 18B active
License	GLM-5.3 License (MIT-like + $10B MaaS clause)	MIT	MIT
Context	1M tokens	1M tokens	1M tokens
Terminal-Bench 2.1	88.2	81.0	—
Weights released	Aug 28 2026	Jun 13 2026	Aug 2026
Min. local setup	24 GB GPU + ~256 GB RAM (2-bit, slow)	Same	Smaller, still >100 GB class

Honest take: if you already run GLM-5.2, the model is better and the hardware bill is identical — the only new line item is a license that is no longer a rubber stamp. If you were never going to self-host a 744B model anyway, nothing here changes that.

---

What actually changed from GLM-5.2 to GLM-5.3?

Post-training, and only post-training. Z.ai announced GLM-5.3 on August 14, 2026 (API first) and confirmed it reuses the GLM-5.2 base model — 744B total parameters, roughly 40B active per forward pass, the same MoE backbone — with all of the gains coming from an expanded post-training and RL phase. The 1M-token context window and 128K output ceiling carry over unchanged.

The open weights did not land with the announcement. They shipped on Hugging Face on August 28, 2026, after a roughly two-week safety review — the first time Z.ai has held weights back for a staged release. That two-week gap matters for one practical reason: it is also when the license changed (next section).

Because the base model is identical, everything operational you know from GLM-5.2 transfers: the GGUF quant sizes are in the same ~400–500 GB class at 4-bit, the chat template family is the same, and a vLLM config that served 5.2 serves 5.3 with a model-path swap. Our GLM-5.2 self-hosting review has the full VRAM table; the numbers have not moved.

Is GLM-5.3 still MIT-licensed? No — read the new clause

GLM-5.2 shipped under plain MIT. GLM-5.3 does not. It ships under a custom “GLM-5.3 License” that keeps MIT-equivalent grants for most users but adds one condition: a licensee that (a) operates a Model-as-a-Service business — defined as giving third parties access to inference or fine-tuning with meaningful control over inputs, parameters, or training data — and (b) has aggregate revenue, including affiliates, above $10 billion over any consecutive 12 months, must pass Z.ai’s security review before any commercial use of the weights or derivatives (reported consistently by The New Stack and noze.it, Aug–Sep 2026).

What this means in practice, as of October 2026:

• A home lab, a startup, or a mid-size company self-hosting for internal use: unaffected. You are not a MaaS provider, and you are nowhere near the revenue threshold.
• A company reselling GLM-5.3 inference (an OpenRouter-style aggregator, a cloud offering fine-tuning): only affected above $10B revenue — this clause is aimed at hyperscalers, not at you.
• Anyone whose compliance process says “OSI-approved licenses only”: affected immediately. The GLM-5.3 License is not OSI-approved and will not pass an automated license scan the way MIT did. GLM-5.2’s weights remain MIT and remain downloadable — that is your fallback.

The pattern is worth naming: this is the same move Meta made with the Llama Community License’s 700M-MAU clause — a restriction designed to bind competitors, which nonetheless removes the model from the “no legal review needed” bucket. Note also that GLM-5.3-Flash, the smaller 320B/18B multimodal sibling, stays plain MIT. Z.ai is drawing the line at the flagship only. For how these conditional licenses compare across vendors, see our open-weight licenses that block commercial use breakdown.

How much VRAM does GLM-5.3 need?

The same as GLM-5.2, because it is the same base model: at BF16 the weights are roughly 1.5 TB; production FP8 serving needs an 8×H200-class node (~860 GB VRAM plus KV cache headroom at long context). The 4-bit GGUF tier sits in the ~450–480 GB range — multi-GPU server territory, not a workstation.

The floor for “it technically runs locally” is unchanged too: a 24 GB GPU plus ~256 GB of system RAM running Unsloth’s 2-bit dynamic quant with MoE expert offload. It works, it is slow (expect painful time-to-first-token on long prompts), and 2-bit quantization measurably degrades code quality — which partially cancels the reason you wanted 5.3 over 5.2 in the first place.

One real constraint we hit while verifying this article: as of October 7, 2026, huggingface.co was unreachable from our build environment, so the Unsloth GGUF variant list for unsloth/GLM-5.3-GGUF (which search snapshots confirm exists, with a UD-Q4_K_XL tag) could not be byte-checked against the 5.2 sizes. Since both models share the 744B base, treat GLM-5.2’s published quant sizes as the planning numbers and verify the exact file size on the Hugging Face page before you commit a download — a ~450 GB pull on a metered connection is not a casual mistake.

How do you run GLM-5.3 with Ollama or vLLM?

For the GGUF path, Ollama pulls the Unsloth quant directly from Hugging Face:

ollama run hf.co/unsloth/GLM-5.3-GGUF:UD-Q4_K_XL
# pulling manifest... (expect a multi-hundred-GB download at 4-bit)
# >>> Send a message (/? for help)

Check that inference is actually local before you send anything sensitive:

ollama ps
# NAME                                    SIZE      PROCESSOR    UNTIL
# hf.co/unsloth/GLM-5.3-GGUF:UD-Q4_K_XL   ~450 GB   CPU/GPU      ...

The trap from GLM-5.2 still applies: Ollama’s short glm tags include :cloud variants that route inference to Ollama’s servers instead of your GPU. A cloud model shows zero local VRAM in ollama ps. If privacy is why you are self-hosting, pull the explicit hf.co/... path every time — and if your instance is reachable from outside your LAN, read our exposed Ollama instances guide first.

For vLLM, GLM-5.3 uses the same model class as GLM-5.2, so the serve command is a path swap:

vllm serve zai-org/GLM-5.3 --tensor-parallel-size 8 --tool-call-parser glm

Verify the chat template against the model card before production use. Z.ai has not announced a template break between 5.2 and 5.3, and the shared base makes one unlikely, but “unlikely” is not “confirmed” — run one tool-call round-trip as a smoke test before pointing an agent framework at it.

Is GLM-5.3 actually better? The benchmark picture

On agentic coding, clearly yes. Per the public comparison trackers (benchlm.ai, gradually.ai, checked October 2026):

Benchmark	GLM-5.3	GLM-5.2
Terminal-Bench 2.1	88.2	81.0
Terminal-Bench 3.0	28.3	4.6
SWE-bench Verified	95.4	82.8

Two caveats before you quote these. First, the Terminal-Bench 3.0 numbers look shocking (28.3 vs 4.6) because 3.0 is a far harder benchmark released after GLM-5.2’s training — a 6× relative jump on it is consistent with “post-trained specifically for current agentic evals,” which cuts both ways. Second, these are aggregator-listed scores, not numbers we reproduced; treat single-benchmark deltas as directional. The honest summary: GLM-5.3 is a meaningful agentic upgrade on the same brain, not a new brain.

When NOT to self-host GLM-5.3

• Your compliance policy requires OSI-approved licenses. GLM-5.3’s custom license fails that filter automatically. Stay on MIT-licensed GLM-5.2, or use GLM-5.3 via API where the weights license is irrelevant.
• You have less than ~256 GB of system RAM. There is no configuration where this model fits. A 24–32 GB GPU runs excellent dense models (Qwen3.6 32B class) natively instead.
• You need interactive latency on a budget. RAM-offloaded 2-bit MoE inference is batch-job territory. For live pair-programming, the API wins on both speed and quality.
• You already run GLM-5.2 and your workloads are not agentic. For plain chat and RAG summarization the two models are near-identical; a fresh ~450 GB download buys you little. The broader framework is in when NOT to self-host AI.

If you want GLM-5.3 as a coding backend rather than an infrastructure project, the BYOK route through Cline or Cursor against the Z.ai API is covered from the coding-tool side at aicoderscope.com’s GLM backend guide — the 5.2 setup there transfers to 5.3 unchanged.

What to actually buy

Prices as of October 2026, verified against the comparison above:

Your situation	The move	Price	Where
Want GLM-5.3 quality with zero ops	Z.ai API (weights license irrelevant)	per token	z.ai
Want to benchmark the 4-bit GGUF before buying anything	Rent a multi-GPU box by the hour	from $0.14/hr per 4090-class GPU	Vast.ai
Building a general local-LLM box (not for 744B models)	Used RTX 3090 24 GB	~$1,250 (Sep 2026 street)	Check price
Need OSI-clean weights today	GLM-5.2 (MIT, still published)	hardware only	Hugging Face

The first column is the point: a single consumer GPU is a fine purchase, but not for this model — buy it for the 30B-class models it actually fits, and rent by the hour for anything GLM-sized. GPU and RAM pricing for a local box is covered in depth on runaihome.com, and remember that DDR5 prices are still ~4× their 2025 levels, so the 256 GB RAM “budget” path costs more than the GPU.

FAQ

Can I use GLM-5.3 commercially? Almost certainly yes. The GLM-5.3 License grants MIT-equivalent rights unless you are a Model-as-a-Service provider with over $10B in aggregate 12-month revenue, in which case you need Z.ai’s security review first. Internal commercial use, fine-tuning, and products built on the model are unrestricted below that threshold — but the license is custom, not OSI-approved, so run it past legal if your policy requires standard licenses.

Should I upgrade from GLM-5.2 to GLM-5.3? If your workload is agentic coding (multi-step tool use, repo-scale editing), yes — the Terminal-Bench 2.1 jump from 81.0 to 88.2 is real headroom, hardware requirements are identical, and your existing vLLM config transfers. If your workload is chat or RAG, or your compliance requires MIT, stay on 5.2.

What is GLM-5.3-Flash and can I run it at home? GLM-5.3-Flash is the 320B-total/18B-active multimodal sibling, and unlike the flagship it is plain MIT. It is less than half the flagship’s size but still a >100 GB-class download at useful quants — easier to rent, still out of reach for a single 24 GB card at good quality.

Sources

• The New Stack — Z.ai GLM weights license analysis (license clause text, Sep 2026)
• noze.it — GLM-5.3: the weights, the licence and the benchmark method
• benchlm.ai — GLM-5.2 vs GLM-5.3 comparison (benchmark scores)
• intuitionlabs.ai — GLM-5.3 vs GLM-5.3-Flash deployment (Flash specs)
• Unsloth GLM-5.3 GGUF on Hugging Face (quant tags)

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe