TL;DR: Z.ai published the full GLM-5.3 weights on August 28, 2026 — two weeks late, as promised, and under a custom “GLM-5.3 License” instead of GLM-5.2’s MIT. The hardware math is exactly what we predicted: same 744B/40B-active base, smallest usable quant 239GB. For home labs the license change is a non-event and the weights are still un-runnable; GLM-5.3-Flash remains the one to actually download.

	GLM-5.3 full weights	GLM-5.3-Flash (Aug 26)	GLM-5.3 via API
Best for	256GB+ RAM builds, patience required	128GB unified-memory boxes, today	Everyone else
Memory floor	239GB (2-bit GGUF) + KV	93GB (1-bit GGUF)	$0
License	Custom GLM-5.3 License	MIT	n/a (terms of service)
The catch	3–9 tok/s on anything you can afford	llama.cpp merge still in flight	Your code leaves your machine

Honest take: The weights dropped, the license got a hyperscaler clause your home lab will never trigger, and not one gigabyte of the hardware wall moved. Unless you already own a 256GB machine, treat this release as a license story and a security story — and keep using the API or Flash.

Two weeks ago, when GLM-5.3 launched API-only, we told you to watch for the weights around August 28 and promised the hardware math wouldn’t change. Both calls landed: the weights went up on Hugging Face on August 28, and the quant sizes match GLM-5.2’s almost to the gigabyte. What nobody predicted was the license. For the first time in the GLM-5 series, the flagship did not ship under MIT — and the reason connects directly to the 2,436 security vulnerabilities Z.ai says this model found during its two-week quarantine.

Here’s what actually landed, what the new license does and doesn’t restrict, and what — if anything — your hardware should do about it.

What landed on August 28

Z.ai published the full-precision GLM-5.3 repos under its zai-org Hugging Face account, in BF16 and FP8 flavors. At BF16 the weights alone run near 1.5TB; FP8 roughly halves that — before you allocate a single byte of KV cache. Nothing about the shape is new: GLM-5.3 reuses the GLM-5.2 base, a ~744B-total mixture-of-experts model with ~40B active parameters per token and a 1M-token context window with 128K max output. Every improvement came from post-training.

Unsloth’s dynamic GGUFs followed within hours, true to the pattern set by every GLM release since June. The ladder that matters:

Quant	Disk size	Minimum memory (RAM + VRAM)	Realistic speed
UD-IQ1_S (1-bit dynamic)	~217GB	~230GB	3–6 tok/s
UD-IQ2_M (2-bit dynamic)	239GB	245GB+	3–9 tok/s
UD-Q4_K_XL (4-bit dynamic)	~420GB	~450GB	Server territory
FP8 (official)	~750GB	8× datacenter GPUs	Rent it

The 2-bit number is the one to memorize: 239GB on disk, 245GB+ of combined RAM and VRAM to run — identical to GLM-5.2’s footprint, which is what an unchanged base gets you. The speed estimates carry over from measured GLM-5.2 experience too: a 256GB Mac Studio or a 256GB DDR5 CPU-offload build with one GPU produced 3–9 tok/s on the 2-bit quant, and 40B active parameters per token means that ceiling is set by memory bandwidth, not by anything Z.ai’s post-training touched.

One genuinely useful day-one difference from Flash: the flagship loads in stock llama.cpp. Because GLM-5.3 shares GLM-5.2’s architecture — which merged into mainline back in June — Unsloth’s guide points straight at a standard build, no special branch. Download and run it like this:

# grab the 2-bit shards (6 files, 239GB total)
huggingface-cli download unsloth/GLM-5.3-GGUF \
  --include "UD-IQ2_M/*" --local-dir ./glm-5.3

# run with MoE expert layers offloaded to system RAM
./llama-server -m ./glm-5.3/UD-IQ2_M/GLM-5.3-UD-IQ2_M-00001-of-00006.gguf \
  -ngl 99 -ot ".ffn_.*_exps.=CPU" -c 8192 --flash-attn

Watch the startup log for offloaded 99/99 layers to GPU — with the -ot regex, that means attention and shared layers live in VRAM while the bulky expert weights stay in system RAM, which is the only sane split for a 239GB model on consumer hardware.

Contrast that with GLM-5.3-Flash, whose new glm5_next architecture was still waiting on its mainline llama.cpp merge when we covered it on August 28. That merge is now in flight — two competing pull requests landed within a day of each other, including one from Unsloth’s Daniel Han — so the Flash catch we flagged should resolve within days. If you pulled a Flash GGUF and hit unknown model architecture: 'glm5next', that’s the fix: update llama.cpp once the merge lands, or use Unsloth’s branch today. Our unknown-architecture troubleshooting guide covers the general pattern.

The license: what “not MIT” actually means for you

GLM-5.2 shipped MIT — full stop, no strings. GLM-5.3 ships under a custom GLM-5.3 License, and it’s worth being precise about what changed, because “Z.ai abandoned open source” takes are already circulating and they’re mostly wrong.

The license still permits use, modification, distribution, sublicensing, sale, deployment, fine-tuning, and derivative works, with the usual keep-the-notice condition. What’s new is one clause: any company whose aggregate revenue exceeds $10 billion over any 12 consecutive months must pass Z.ai’s security review before using the model or its derivatives commercially. The clause is aimed at hyperscalers and the larger neoclouds — model-as-a-service operators who would rehost GLM-5.3 for third parties.

Run the test on yourself: does your home lab, your consultancy, or your startup clear $10 billion in trailing-12-month revenue? Then nothing in this license touches you. You can fine-tune it, ship products on it, and sell access to it. The practical differences from MIT for a home-lab user are close to zero — but it is not OSI-approved open source, and if your employer has a strict license-allowlist policy, “custom license with a revenue-triggered review clause” may not clear legal the way MIT did. That’s a real cost for some readers, and pretending otherwise would be the kind of rounding-up this site exists to avoid.

Worth noting: Z.ai hedged its own bet here. GLM-5.3-Flash shipped two days earlier under clean MIT. The 320B model most home labs can actually run kept the permissive license; the 744B flagship with the exploit-writing skills got the guardrails.

Why the weights sat in quarantine for two weeks

The two-week delay wasn’t theater. During the hold, Z.ai says it ran GLM-5.3 against real open-source codebases, and the numbers it published are startling: 2,436 vulnerabilities surfaced across 269 open-source projects, 1,097 of them rated critical or high severity, including bugs in Linux, WebKit, and FreeBSD. The oldest defect dated to 1981 — a mean of 26.6 years between a bug being introduced and this model finding it. As of the weights release, 53 of those vulnerabilities have public CVEs; 2,383 remain under embargo while maintainers patch. One of the reported finds was a serious vulnerability in Cursor, the AI code editor.

The benchmark behind the story is CyberGym, where Z.ai reports 84.5% — a state-of-the-art result it says edges past Claude Mythos 5 and GPT-5.6 Sol at vulnerability discovery. The usual caveat applies double here: every one of these numbers is vendor-reported, no independent lab has re-run them, and a security benchmark is exactly the kind of thing a vendor has incentive to frame generously. Treat 84.5% as Z.ai’s best case.

But the direction is credible, and it explains the license. An openly downloadable model that is genuinely good at finding and chaining exploits is a dual-use release in a way a coding model isn’t. The security-review clause, the two-week hold, and the coordinated-disclosure ledger are Z.ai’s answer to a question that will keep getting harder for every open-weights lab: how do you ship capability like this to everyone without shipping it to anyone? If you run an exposed local AI stack, this cuts both ways — the same class of model that finds bugs in WebKit can probe your Open WebUI instance, which is why we hardened our own Ollama setups last week.

The hardware decision, tier by tier

Nothing in this release changes what your money should do, but the release makes the map worth restating — because the answer now differs sharply by memory tier.

24GB (used RTX 3090, RTX 4090): GLM-5.3 is not for you, in any quant, at any patience level. A used RTX 3090 — $1,264 average in August 2026, $1,201–$1,299 fair range across 366 listings — remains the value play for the 27B–35B class. Qwen3.6-35B-A3B at 50–65 tok/s is a better daily driver than a 3 tok/s frontier model would ever be, and the 24GB tier guide has the full menu. Pocket the GLM-5.3 API for the hard problems.

128GB unified memory (Strix Halo boxes, 128GB Macs): GLM-5.3-Flash’s 93GB 1-bit quant is the most capable MIT-licensed thing you can physically load, at an estimated 10–20 tok/s on a Strix Halo machine. The full GLM-5.3 doesn’t fit at any quant. See what else earns a slot at this tier.

256GB+ (Mac Studio 256GB, DDR5 CPU-offload server): You can genuinely run the 239GB 2-bit quant today, in stock llama.cpp, at 3–9 tok/s. That’s usable for overnight batch work, unattended agents, and privacy-mandated analysis — not for interactive coding. The 2-bit quality tax on GLM-5.3 is also unmeasured as of this writing; nobody has published perplexity or benchmark deltas for the fresh GGUFs. If your workload justifies a Mac Studio M3 Ultra or a 256GB EPYC build, it justified it before Friday too.

The $16,000 question: the single-card path to big-model VRAM, the RTX PRO 6000, was repriced to $16,000 on August 13 — and even four of them (384GB, ~$64,000) only clear the 2-bit quant with room for context. The $52K four-card GLM-5.2 build we profiled in July hit 80 tok/s using a REAP expert-pruned NVFP4 variant; no equivalent pruned GLM-5.3 exists yet, though the identical base means the same treatment should apply when the community gets to it. Until then, renting is strictly cheaper: an 8× H100 pod on RunPod runs the FP8 checkpoint for an afternoon of evaluation at a tiny fraction of one card’s price. The rent-vs-buy math hasn’t moved.

Everyone: the API stayed at $1.40 per million input tokens, $4.40 output, $0.26 cached. At those prices, $64,000 of GPUs buys you roughly 14 billion output tokens — decades of personal usage. The hardware case for GLM-5.3 is privacy and always-on agents, never economics.

If you’re wiring any GLM-5.3 variant into a coding workflow, our sister site covers local and API coding-tool backends, and aifoss.dev tracks the FOSS self-hosting stack around these releases.

FAQ

Is GLM-5.3 still open source? It’s open-weight under a custom license, not OSI open source. For individuals and any business under $10B in trailing-12-month revenue, the permissions match MIT in practice: use, modify, fine-tune, distribute, sell. Above that line, Z.ai requires a security review before commercial use.

Did the GGUF sizes change from GLM-5.2? No — same base, same footprint. The 2-bit dynamic quant is 239GB for both (Unsloth lists 245GB+ combined memory as the floor for GLM-5.3’s UD-IQ2_M). GLM-5.2 hardware planning transfers one-for-one.

Does the flagship run in Ollama or LM Studio? The architecture is supported (it’s GLM-5.2’s), so the blocker is size, not software: you need ~245GB of combined memory before any of those front-ends matter. Flash is the opposite case — small enough for real machines but waiting on the glm5next llama.cpp merge to propagate to Ollama and LM Studio.

Should I download the weights now in case the license changes? The license travels with the copy you downloaded — a future change can’t retroactively revoke what you have. If you have 239GB of disk and any plausible future 256GB machine, mirroring the 2-bit GGUF is cheap insurance. But there’s no signal Z.ai intends to pull them.

Is GLM-5.3 better than GLM-5.3-Flash at coding? Z.ai’s own numbers say yes — the flagship leads every benchmark it publishes, and it reports a 50% jump over GLM-5.2 on its internal code bench. But Flash runs on hardware that costs less than a used car, and a model you can run beats a model you can’t. For a 128GB box, Flash is the right answer even at a quality discount.

Recommended Gear

• NVIDIA RTX 3090 (used) — still the 24GB value king at ~$1,264; runs the 27B–35B class that actually makes sense at home
• Apple Mac Studio M3 Ultra — the 256GB configuration is the quietest path to running GLM-5.3’s 2-bit quant at all

Sources

• Z.ai’s GLM-5.3 goes open weight, but its new license aims at hyperscalers — The New Stack
• unsloth/GLM-5.3-GGUF — Hugging Face
• GLM-5.3: How to Run Locally — Unsloth Documentation
• Preparing GLM-5.3 for Open Release: A Responsible Path to Cyber Defense — Z.ai
• GLM-5.3 — Overview — Z.AI Developer Documentation
• GLM-5.3 Open Weights: Release Date, License, Bug Ledger — Modemguides
• GLM-5.3’s Open-Weight Reality Check: The Two-Week Delay, Cybersecurity Leap, and 2,436-Vulnerability Claim — Kingy AI
• GLM-5.3 found a serious vulnerability in Cursor — VentureBeat
• GLM-5.3 Weights Are Out — But Running Them Takes Eight GPUs — Kingy AI
• model: add GLM-5-Next (GLM-5.3-Flash) — llama.cpp PR #27754
• NVIDIA RTX 3090 used price tracker — ResalePrices
• GLM-5.3 Is Out: What Shipped and What’s Still Staged — Evolink

Last updated August 29, 2026. Prices and specs change; verify current rates before purchasing.