GLM-5

Z.ai's February 2026 MIT-licensed 744B-A40B flagship that moved GLM from coding to engineering; it has three newer successors.

In short. GLM-5 is a very large text model from Z.ai for coding and long agent tasks. Its license is plain MIT. It needs a GPU server or a machine with very large memory, and Z.ai has since released GLM-5.1, 5.2 and 5.3, so it now mainly matters as the base of that line.

Summary

Z.ai released GLM-5 on 12 February 2026 for complex systems engineering and long-horizon agent tasks. According to its card, GLM-5 grows from GLM-4.5's 355 billion parameters (32 billion active) to 744 billion (40 billion active) and from 23 to 28.5 trillion tokens of pre-training data. Hugging Face's file metadata counts about 754 billion parameters. It integrates DeepSeek Sparse Attention to cut deployment cost while keeping long-context ability, and it was post-trained with slime, an asynchronous reinforcement-learning system Z.ai built to raise training throughput. The configuration lists 78 layers, 256 routed experts with 8 used per token plus one shared expert, and 202,752 positions. The technical report is arXiv 2602.15763.

In Z.ai's own table GLM-5 scores 77.8 on SWE-bench Verified (Claude Opus 4.5 80.9, GPT-5.2 at xhigh effort 80.0, Kimi K2.5 76.8, GLM-4.7 73.8), 73.3 on SWE-bench Multilingual, 56.2 on Terminal-Bench 2.0 with Terminus-2 (60.7 on Z.ai's verified version of the dataset), 62.0 on BrowseComp, 89.7 on τ²-Bench, 67.8 on MCP-Atlas, 30.5 on Humanity's Last Exam without tools and 50.4 with tools. Z.ai calls it the best of the open models on reasoning, coding and agentic tasks at release. These are self-reported scores, and newer models have since overtaken them, including Z.ai's own.

The BF16 repository is 1,507.8 GB, an FP8 build of 756.2 GB is published as GLM-5-FP8, and by our estimate a 4-bit build needs about 520 GB of memory. The card lists vLLM 0.19.0 or newer, SGLang 0.5.10 or newer, KTransformers, Transformers and xLLM for Ascend NPUs, and its vLLM example uses 8-way tensor parallelism with multi-token-prediction speculative decoding. GLM-5.1 followed in April, GLM-5.2 in June and GLM-5.3 in August 2026. In the 30 days to 11 November 2026, Hugging Face counted about 1,319,000 downloads of the main repository and about 49,000 of the FP8 build, and the main count may include mirrors and automated pulls.

Good for

  • Self-hosted coding and engineering agents on a GPU server
  • Teams that want a plain MIT license with no regional or revenue conditions
  • Research on DeepSeek Sparse Attention and asynchronous reinforcement learning
  • Baselines when comparing the GLM-5 line with its successors

Watch out for

  • GLM-5.1, 5.2 and 5.3 are newer and score higher on Z.ai's main coding and agent benchmarks, so check them before starting a new deployment
  • It needs a multi-GPU node even in FP8, or a machine with roughly 512 GB of fast memory for a 4-bit build
  • It reads text only, and its 202,752-position context is far below GLM-5.2's 1M tokens
  • All benchmark numbers come from Z.ai and use its own harness settings

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the GLM-5 (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
GLM-5 (BF16) BF16 754B 40B active 1.73 TB 867 GB 520 GB GPU server
GLM-5-FP8 FP8 754B 40B active 1.73 TB 867 GB 520 GB GPU server

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.

More about MIT

Frequently asked questions

What is GLM-5?
GLM-5 is a very large text model from Z.ai for coding and long agent tasks. Its license is plain MIT. It needs a GPU server or a machine with very large memory, and Z.ai has since released GLM-5.1, 5.2 and 5.3, so it now mainly matters as the base of that line.
Can I use GLM-5 commercially?
Yes. GLM-5 is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is GLM-5 open source?
GLM-5 is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run GLM-5 on my own computer?
Not on a desktop. The smallest variant, GLM-5 (BF16), needs about 520 GB at 4-bit precision, which means a GPU server with several data-center cards.
When was GLM-5 released?
GLM-5 was released by Z.ai (Zhipu AI) on February 12, 2026.
How long a context window does GLM-5 have?
GLM-5 supports a context window of up to 198K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
GLM-5.1 Z.ai (Zhipu AI) 754B 40B active 198K MIT (Permissive) 520 GB GPU server
GLM-5.2 Z.ai (Zhipu AI) 753B 40B active 1M MIT (Permissive) 520 GB GPU server
GLM-5.3 Z.ai (Zhipu AI) 753B 40B active 1M GLM-5.3 License (Conditional) 520 GB GPU server
DeepSeek-V3.2 DeepSeek 685B 37B active 160K MIT (Permissive) 473 GB Big-memory desktop