GLM-5.3
Z.ai's ~753B parameter flagship, rebuilt through post-training on the GLM-5.2 base, and GLM-5.3-Flash, a 320B MIT-licensed sibling that fits a single 512GB Mac Studio on Faraday's Scale tier.
Model Story
GLM-5.3 is the latest flagship from Z.ai (formerly Zhipu), the team behind the General Language Model family that started at Tsinghua University. Z.ai launched it on its API on August 14, 2026 and published the weights on Hugging Face on August 28, 2026. It follows GLM-5.2 (June 16, 2026) and GLM-5.1, and it keeps the same ~753B total / ~40B active Mixture-of-Experts base as GLM-5.2. All of the gains come from post-training. The context window is 1M tokens.
Those gains are large on Z.ai's own numbers. GLM-5.3 scores 66.9 on DeepSWE v1.1, up from GLM-5.2's 46.2, and 28.3 on Terminal-Bench 3.0, up from 4.6. Independently, Artificial Analysis scores GLM-5.3 at 45 on its Intelligence Index v4.3, second among open-weight models behind MiMo-V2.6-Pro (46).
One thing did get worse for buyers: the license. GLM-5.2 was MIT. GLM-5.3 uses a bespoke "glm-5.3" license whose main restriction is that model-as-a-service businesses above $10B in annual revenue must pass Z.ai's security review. Few organizations come near that threshold, but it is no longer the "no review needed" MIT license GLM was known for.
Alongside the flagship, Z.ai released GLM-5.3-Flash: a smaller 320B model under a plain MIT license. GLM-5.3-Flash is the GLM model Faraday Machines ships on standard tiers, and it has its own section below.
Key Specifications
| Developer | Z.ai (formerly Zhipu) |
|---|---|
| Release Date | August 14, 2026 (API); weights published August 28, 2026 |
| Architecture | Sparse Mixture-of-Experts (MoE), same base as GLM-5.2 |
| Total Parameters | ~753 billion |
| Active Parameters | ~40 billion per token |
| Context Window | 1M tokens |
| License | Bespoke "glm-5.3" license (not MIT) |
| Weights | Hugging Face |
| Size at 4-bit | Roughly 410 GB |
| Size at 1- to 2-bit | About 223–239 GB (Unsloth builds) |
| Faraday Fit | Scale (1 Mac Studio M5 Ultra, 512GB) at 1- to 2-bit; larger setups at 4-bit (talk to us) |
API Pricing
Z.ai list price for GLM-5.3 as reported by API aggregators, October 2026. On-premises deployment on Faraday Machines eliminates per-token costs entirely; you pay only the monthly plan and electricity.
Benchmarks
The top row is vendor-reported by Z.ai. Z.ai also reports 31.4% on its own Z.ai Code Bench, ahead of Claude Opus 4.8 (29.5%) while using under half the output tokens but behind Fable 5 (39.5%), and lists GPT-5.6 Sol at 34.6 on Terminal-Bench 3.0. The bottom row is independent, from the Artificial Analysis Intelligence Index v4.3. Under the earlier v4.1.1 index, GLM-5.3 (max) scored 60 at $0.68 per task; those older scores are not comparable with v4.3. GLM-5.3 had no standardized SWE-bench Pro entry as of late September 2026.
GLM-5.3-Flash: The Version Faraday Ships
GLM-5.3-Flash is the practical GLM for on-premises work. It is a 320B total / 18B active Mixture-of-Experts model with a 1M-token context, natively multimodal, and released under a plain MIT license, so it carries none of the flagship's license conditions.
Artificial Analysis scores GLM-5.3-Flash at 42 on its Intelligence Index v4.3 at $0.25 per task. That is the same score as GPT-5.6 Terra (42 at $1.40 per task) and Claude Opus 5.5 at low effort (42 at $0.55 per task), from a model you can run on two Mac Studios.
| Total Parameters | 320 billion |
|---|---|
| Active Parameters | 18 billion per token |
| Context Window | 1M tokens |
| Multimodal | Native multimodal input |
| License | MIT (full commercial freedom) |
| Size at 4-bit | Roughly 180 GB |
| Size at 1- to 2-bit | About 93–109 GB (Unsloth builds) |
| Faraday Fit | Scale (1 Mac Studio M5 Ultra, 512GB) at 4-bit; Growth (1 Mac Studio M5 Ultra, 256GB) at 1- to 2-bit |
On-Premises Deployment
GLM-5.3-Flash is the GLM model included on Faraday's standard tiers. At 4-bit quantization it is roughly 180 GB, which fits the Scale tier (one Mac Studio M5 Ultra, 512GB of unified memory) with room left for working memory. Faraday recommends it for general business intelligence, document analysis and multimodal input. Unsloth also publishes 1-bit (about 93 GB) and 2-bit (UD-Q2_K_XL, about 109 GB) builds that fit the Growth tier (one Mac Studio M5 Ultra, 256GB). Unsloth says they retain about 71% and 78% accuracy; that is Unsloth's own measure of how closely each build matches the full model, not a benchmark score. Lower-bit builds trade some quality, most noticeably on long or precise work.
The full GLM-5.3 runs on Scale as a 1- to 2-bit build; at 4-bit it needs a larger setup. At 4-bit quantization its weights are roughly 410 GB, too much for the 512GB in our single-machine Scale tier once working memory is counted, so a 4-bit deployment is sized by quote. Unsloth's 1-bit build is about 223 GB and its 2-bit build (UD-IQ2_M) about 239 GB, so both fit Scale with room for working memory. Unsloth says they retain about 76% and 81% accuracy, again its own comparison with the full model rather than a benchmark score, so expect the quality trade-off to show most on long or precise work. Organizations that want the flagship's extra coding and agentic performance, and are comfortable with the bespoke license, can size a deployment with us.
Frequently Asked Questions
Can I run GLM-5.3 on-premises?
Yes. GLM-5.3-Flash, the 320B MIT-licensed sibling, runs on Faraday Machines' Scale tier, a single Mac Studio M5 Ultra: at 4-bit quantization it needs roughly 180 GB of the 512GB of unified memory. Unsloth's 1- and 2-bit builds of GLM-5.3-Flash, about 93–109 GB, fit the Growth tier, a single Mac Studio M5 Ultra with 256GB. The full GLM-5.3 needs roughly 410 GB at 4-bit, too much for Scale's 512GB once working memory is counted, so a 4-bit deployment is a larger setup we size with you. Unsloth's 1- and 2-bit builds of the full model, about 223–239 GB, fit Scale. Lower-bit builds trade some quality, most noticeably on long or precise work.
Is GLM-5.3 still MIT licensed?
No. GLM-5.3 uses a bespoke glm-5.3 license instead of MIT. Its main restriction is that model-as-a-service businesses above $10B in annual revenue must pass Z.ai's security review. GLM-5.2 was MIT, and GLM-5.3-Flash, the smaller 320B model Z.ai released alongside GLM-5.3, still ships under a plain MIT license with full commercial freedom.
How much does GLM-5.3 cost?
Z.ai's list price for GLM-5.3 is $1.40 per 1M input tokens and $4.40 per 1M output tokens, as reported by API aggregators in October 2026. Artificial Analysis measures GLM-5.3-Flash at $0.25 per Intelligence Index task. Running GLM-5.3-Flash on-premises on a Faraday Machines Scale tier ($999 US a month on a 36-month plan, including hardware and support) removes per-token costs entirely.
How does GLM-5.3-Flash compare to GPT-5.6 Terra and Claude Opus 5.5?
GLM-5.3-Flash matches both on the Artificial Analysis Intelligence Index v4.3 at a lower cost per task. It scores 42 at $0.25 per task, the same score as GPT-5.6 Terra (42 at $1.40 per task) and Claude Opus 5.5 at low effort (42 at $0.55 per task). The full GLM-5.3 scores 45, second among open-weight models behind MiMo-V2.6-Pro at 46.
References
[1] AI News. (2026). "Zhipu GLM-5.3 benchmarks explained." August 2026. Available at: artificialintelligence-news.com
[2] Digital Applied. (2026). "GLM-5.3 weights ship under a bespoke license, not MIT." August 2026. Available at: digitalapplied.com
[3] Artificial Analysis. (2026). "GLM-5.3" model page. Available at: artificialanalysis.ai
[4] Artificial Analysis. (2026). "GLM-5.3-Flash" model page. Available at: artificialanalysis.ai
[5] Artificial Analysis. (2026). "Artificial Analysis Intelligence Index v4.3." September 2026. Available at: artificialanalysis.ai
[6] Unsloth. (2026). "GLM-5.3" documentation (dynamic GGUF builds). Available at: unsloth.ai
[7] Unsloth. (2026). "GLM-5.3-Flash" documentation (dynamic GGUF builds). Available at: unsloth.ai
Explore Other Models
See how every model compares on price, benchmarks and hardware fit in the AI Model Comparison.
Run GLM-5.3-Flash on Hardware You Control
A single Faraday Scale Mac Studio (512GB) runs MIT-licensed GLM-5.3-Flash on-premises at 4-bit. No per-token charges, no license review, full data sovereignty. Need the full GLM-5.3? It runs on Scale as a 1- to 2-bit build; at 4-bit it needs a larger setup, so talk to us.
Get a Fixed-Cost Quote