What Do You Need to Run GLM Locally? VRAM by Version
GLM currently has 5 versions you can deploy locally. The smallest, GLM 4.7 Flash, needs 20GB of VRAM at 4-bit; the largest, GLM 5.3, is an MoE model and needs 14GB of VRAM plus about 457GB of RAM with MoE offloading. The cheapest device that gets one running is the Tesla V100 32GB, at about $895.
| Version | Released | Context | License | ||
|---|---|---|---|---|---|
| GLM 5.3MoE | Aug 28, 2026 | 744B | 40B | 1M | GLM-5.3 License |
| GLM 5.3 FlashMoE | Aug 26, 2026 | 320B | 18B | 1M | MIT |
| GLM 5.2MoE | Jun 16, 2026 | 744B | 40B | 1M | MIT |
| GLM 4.7 FlashMoE | Jan 19, 2026 | 30B | 3B | 198K | MIT |
| GLM 4.5 AirMoE | Jul 28, 2025 | 106B | 12B | 128K | MIT |
About GLM
GLM is the large model family from Zhipu (Z.ai). Zhipu was founded in Beijing in 2019, and its technology came out of the Knowledge Engineering Group (KEG) in Tsinghua University's Department of Computer Science. It has used the Z.ai brand internationally since July 2025 and listed on the Hong Kong Stock Exchange in January 2026. Since open-sourcing the Chinese-English bilingual GLM-130B in 2022, it has kept releasing open-weight models, with a recent focus on agents and coding.
Its best-known early model was ChatGLM-6B, open-sourced in March 2023, which earned more than 40,000 stars on GitHub. The main line later moved to MoE. GLM 4.5, in July 2025, supports both thinking and non-thinking modes; GLM 5, in February 2026, scaled up to 744B parameters and introduced DeepSeek Sparse Attention (DSA); GLM 5.2, in June, reached a 1-million-token context; and GLM 5.3 Flash, in August, was the first in the series to mix sparse and linear attention. On licensing, most open weights since April 2025 have used MIT. GLM 5.3 switched to a custom license under which model service providers with annual revenue above 10 billion US dollars must pass a safety review before commercial use.
- VendorZ.ai / Zhipu AI
- Versions5
- Parameters30B – 744B
- Latest versionGLM 5.3
How much VRAM do GLM models need at 2-bit?
2-bit, also called Q2, covers common GGUF files such as Q2_K and UD-Q2_K_XL. They are the smallest and lose the most quality, which makes them a way to get a model running first when VRAM is tight. Each row in the table below is a GLM version and each column is a context length; each cell is the total VRAM needed to deploy it at Q2 (weights + KV cache + 1.1GB runtime overhead).
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| GLM 5.3744B (40B active) | 253.9GGUF | 257.9 | 266.8 | 278.5 | 349.2 |
| GLM 5.3 Flash320B (18B active) | 108.7GGUF | 110.1 | 111.3 | 112.7 | 121.6 |
| GLM 5.2744B (40B active) | 253.9GGUF | 257.9 | 266.8 | 278.5 | 349.2 |
| GLM 4.7 Flash30B (3B active) | 11.9GGUF | 14.7 | 20.1 | — | — |
| GLM 4.5 Air106B (12B active) | 47.4GGUF | 54.6 | 73.2 | — | — |
How much VRAM do GLM models need at 4-bit?
4-bit, also called Q4, corresponds to the widely used Q4_K_M files, which balance size and quality and are the default choice for local deployment. Each row in the table below is a GLM version and each column is a context length; each cell is the total VRAM needed to deploy it at Q4.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| GLM 5.3744B (40B active) | 467.3GGUF | 471.3 | 480.2 | 491.9 | 562.6 |
| GLM 5.3 Flash320B (18B active) | 199.7GGUF | 201.1 | 202.3 | 203.7 | 212.6 |
| GLM 5.2744B (40B active) | 465.8GGUF | 469.8 | 478.7 | 490.4 | 561.1 |
| GLM 4.7 Flash30B (3B active) | 18.3GGUF | 21.1 | 26.5 | — | — |
| GLM 4.5 Air106B (12B active) | 73GGUF | 80.2 | 98.8 | — | — |
How much VRAM do GLM models need at 8-bit?
8-bit, also called Q8, corresponds to Q8_0 files, which come closest to the original in quality and are also the largest. Each row in the table below is a GLM version and each column is a context length; each cell is the total VRAM needed to deploy it at Q8.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| GLM 5.3744B (40B active) | 801.4GGUF | 805.4 | 814.3 | 826 | 896.7 |
| GLM 5.3 Flash320B (18B active) | 341GGUF | 342.4 | 343.6 | 345 | 353.9 |
| GLM 5.2744B (40B active) | 801.4GGUF | 805.4 | 814.3 | 826 | 896.7 |
| GLM 4.7 Flash30B (3B active) | 31.8GGUF | 34.6 | 40 | — | — |
| GLM 4.5 Air106B (12B active) | 117.5GGUF | 124.7 | 143.3 | — | — |
FAQ
Is GLM open source and free for commercial use?
Mostly, yes. GLM 4.5 Air, GLM 4.7 Flash, GLM 5.2 and GLM 5.3 Flash all use the standard MIT license: you can use them commercially, modify them and redistribute them, as long as you keep the copyright and license notices. The exception is GLM 5.3, which switched to the GLM-5.3 License, Zhipu's (Z.AI) own. Its terms match MIT except for one added clause: a corporate group that runs a “model as a service” business (letting third parties run inference or fine-tuning through an API or similar means), and whose total revenue over 12 consecutive months exceeds the threshold set in the license, must pass Zhipu's safety review before using the model for any commercial purpose. Using the model only for yourself, or building its capabilities into one feature of a product, doesn't count as model as a service. GLM 5.2, which is the same size, is still MIT and has no such clause.
Can you run GLM with Ollama?
Only GLM 4.7 Flash runs locally. Its tag is glm-4.7-flash, which defaults to q4_K_M, and glm-4.7-flash:q8_0 and glm-4.7-flash:bf16 are also available; it needs Ollama 0.14.3 or later. GLM 5.2, GLM 5.3 and GLM 5.3 Flash have only :cloud tags, which run on Ollama's servers. GLM 4.5 Air has no official tag, so searching for glm4.5:air turns up nothing. To run these models locally, use mainline llama.cpp with unsloth's GGUFs; mainline already supports all of their architectures. GLM 5.2, 5.3 and 5.3 Flash can also run from the official weights on vLLM, which has official recipes for all three.
