Z.ai · 2026-08-28 · major
GLM-5.3 weights go public — Z.ai's 753B coding model lands on Hugging Face
GLM-5.3 is now a public download. Z.ai published the 753B mixture-of-experts weights on Hugging Face in BF16 and FP8, two weeks after the model launched behind the API. The model card lists a custom glm-5.3 license.
Z.ai's 753B GLM-5.3 coding model is now a public download on Hugging Face, in BF16 and FP8.
Quick facts
| Maker | Z.ai (Zhipu AI) |
|---|---|
| Parameters | 753B total, mixture-of-experts |
| Context window | 1M tokens |
| License | glm-5.3 (custom, on the model card) |
| Weight formats | BF16 and FP8 (F8_E4M3) |
| Serving support | SGLang, vLLM, Transformers, KTransformers, Unsloth |
| What's new | Same base model as GLM-5.2 — all gains come from post-training |
Benchmarks
| GLM-5.3 | 28.3% | |
|---|---|---|
| GLM-5.2 | 4.6% | |
| Kimi K3 | 17.4% | |
| Opus 4.8 | 21.1% | |
| Fable 5 | 33.7% | |
| GPT-5.6 Sol | 34.6% |
| GLM-5.3 | 48.2% | |
|---|---|---|
| GLM-5.2 | 26.2% | |
| Kimi K3 | 46.7% | |
| DeepSeek-V4 Pro | 43.2% | |
| Opus 4.8 | 41% | |
| GPT-5.6 Sol | 45.8% |
What is it?
GLM-5.3 weights are now open. Z.ai published the full 753B mixture-of-experts model on Hugging Face as zai-org/GLM-5.3, so the coding and agent model that had been API-only can be run on your own hardware. Both a BF16 build and an FP8 (F8_E4M3) build are in the repository. The model card lists the license as glm-5.3, a custom Z.ai license.
How does it work?
The architecture is a mixture of experts with dynamic sparse attention, tagged glm_moe_dsa in the config, and it handles up to 1M tokens of context. Notably, the base model is the same one GLM-5.2 used — Z.ai says every improvement comes from more post-training rather than a fresh pretraining run. SGLang, vLLM, Transformers, KTransformers and Unsloth are all listed as supported ways to serve it.
Why does it matter?
Self-hosting is now an option for a model that posts frontier-adjacent agentic scores. On Z.ai's own table GLM-5.3 tops CyberGym at 84.5 and AutomationBench at 48.2, ahead of Opus 4.8 and GPT-5.6 Sol on both. Teams running coding agents at volume can weigh their own GPUs against per-token API bills — though the high security scores mean the same weights also lower the cost of offensive tooling.
Who is it for?
teams self-hosting coding and agent models
Frequently asked questions
- Is GLM-5.3 free to download and use?
- GLM-5.3 weights are a public download on Hugging Face at zai-org/GLM-5.3, in both BF16 and FP8 builds. The model card lists the license as glm-5.3 — a custom Z.ai license rather than MIT or Apache-2.0, so read the license file before shipping the weights in a product or redistributing them.
- How does GLM-5.3 compare to GLM-5.2?
- GLM-5.3 reuses the GLM-5.2 base model unchanged, so every reported gain comes from extra post-training. Z.ai's model card puts Terminal Bench 3.0 at 28.3 versus 4.6 for GLM-5.2, ExploitBench at 54.4 versus 24.4, DeepSWE at 66.9 versus 46.2, and GDPval-AA v2 at 1769 versus 1508.
- Does GLM-5.3 beat the closed frontier models?
- GLM-5.3 wins some columns and loses others. Z.ai's table shows GLM-5.3 leading on CyberGym (84.5), AutomationBench (48.2) and GDPval-AA v2 (1769), but trailing GPT-5.6 Sol on Terminal Bench 3.0 (28.3 versus 34.6) and DeepSWE (66.9 versus 72.7). Fable 5 leads ExploitBench at 78.0.
- Where do you get the GLM-5.3 code and paper?
- GLM-5.3 code lives at github.com/zai-org/GLM-5, the Apache-2.0 repository Z.ai uses for the whole GLM-5 line. The research paper, 'GLM-5: from Vibe Coding to Agentic Engineering', is on arXiv as 2602.15763. The weights themselves sit on Hugging Face under zai-org/GLM-5.3.
Try it
zai-org/GLM-5.3 on Hugging Face