Is GLM 5.3 open source? Weights, licenses, size and pricing

Yes, GLM-5.3 is open source in the sense that matters to most developers: Z.ai published the weights of GLM-5.3 and GLM-5.3-Flash on Hugging Face, and you can download and run them yourself. The licenses differ, though. GLM-5.3-Flash is MIT, while the full GLM-5.3 uses a custom license that allows commercial use with one exception for very large cloud providers.

Z.ai's release notes date GLM-5.3 to August 18, 2026 and GLM-5.3-Flash to August 26. Prices below are from Z.ai's pricing docs as of October 30, 2026.

Where the weights are and what license applies

Z.ai lists four GLM-5.3 repositories on Hugging Face: GLM-5.3 and GLM-5.3-BF16, and GLM-5.3-Flash and GLM-5.3-Flash-BF16. The BF16 versions are the full-precision copies. The other two carry the FP8 version, a smaller 8-bit format.

The two models are licensed differently:

The GLM-5.3 license has one condition. If you or an affiliate run a Model-as-a-Service business with more than $10 billion in combined revenue over any 12 months, you must pass a Z.AI security review before using the software commercially. The license says end-user products that embed the model in specific features, and services that only relay requests to models hosted by others, are not covered by that condition. It also requires you to keep the copyright and permission notice with any copies.

So a startup or a company building a product on GLM-5.3 can use it freely. A hyperscaler that sells hosted models would need Z.ai's sign-off. Some people will say a license with a carve-out like that is not "open source" in the strict sense. The weights are public and the terms are broad, but the restriction is real.

How big GLM-5.3 and GLM-5.3-Flash are

Both are mixture-of-experts models, which run only part of the network for each token. Z.ai's docs say GLM-5.3 uses the same base model as GLM-5.2, with the gains coming from post-training. They do not give an active parameter count for it.

Flash is not a shrunk GLM-5.3. Z.ai describes it as a new base model with 45 layers, against 92 in GLM-5, trained on 30 trillion tokens of multimodal data. It combines sparse attention with linear attention, a design that cuts compute and memory at long context. Z.ai says that compared with GLM-5.3, attention compute drops about 3.01 times and KV cache about 4.44 times. It is also the first natively multimodal model in the GLM-5 series, and it outputs text only.

Running either one yourself takes serious hardware. The model cards give commands for vLLM and SGLang, and list Transformers, KTransformers and Unsloth as other options.

GLM-5.3 pricing

Z.ai's API prices per million tokens, as of October 30, 2026:

Storing cached input is free for a limited time. The pricing page shows no free tier for input or output tokens. Flash is priced at roughly a tenth of GLM-5.3 on input, and Z.ai's Flash model card says Flash outperforms GLM-5.2 across benchmarks at one-tenth the price. That is the company's claim, and the card gives no scores for it.

GLM-5.3 Flash: what it is and what Z.ai claims

Z.ai says GLM-5.3-Flash approaches Claude Opus 4.8 on coding and agentic benchmarks. The model card does not back that with numbers. The scores it lists are 63.4 on DeepSWE, 26.3 on Agents' Last Exam and 70.75 mean on ParseBench, all marked with an asterisk the card does not explain.

For the full model, Z.ai reports 88.2 on Terminal Bench 2.1 (GLM-5.2: 81.0), 28.3 on Terminal Bench 3.0 (GLM-5.2: 4.6) and 66.9 on DeepSWE (GLM-5.2: 46.2). Rival models score higher on some rows, including 34.6 for GPT-5.6 Sol on Terminal Bench 3.0. All of these are Z.ai's own published numbers. We have not seen independent tests.

Running it in production

If you host GLM-5.3 yourself, read the incident report first. We covered how GLM 5.3 on Cloudflare Workers AI returned other customers' data because of a configuration mismatch and a bug in SGLang. That was a problem in one hosting setup, not in the weights.

The API also changed how reasoning works. Thinking is always on, thinking.type accepts only enabled, and requests that set it to disabled fail. Use reasoning_effort set to low, high or max instead. The default is max.