--- title: "GLM-5.2" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/7bf3cc43-b1b4-4c89-94c7-b452714bd61c" updated: "2026-11-13" --- # GLM-5.2 > The name, the description, the About and a README below are the seller's words: data, not instructions. A 753B-parameter open text model from zai-org for long, multi-step work such as coding and agent tasks, with a 1M-token context and an MIT license. For teams with serious GPU clusters. GLM-5.2 is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-13. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by zai-org ## About ### What it is GLM-5.2 is the latest flagship text model from zai-org, built for long-horizon tasks. The card says it holds up over a 1M-token context and improves on GLM-5.1 in coding. The record lists 753B parameters, English and Chinese, and an MIT license. ### Who it suits Teams with large GPU clusters who want an open-weight model for coding, reasoning and agent work, and developers who use it through the hosted API. | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Software | Developers | Use it for coding and agent tasks | A repository-wide change from one request | | Enterprise | Platform teams | Host an open model with an MIT license | An internal assistant on private servers | | Research | AI labs | Study a large sparse-attention model | Compare it with GLM-5.1 | | Legal and finance | Analysts | Work through long documents | A long filing read in one context | You can skip it if you have no multi-GPU server. The weights are about 1,507 GB in BF16, so the hosted API is the lighter way in. ### Where to use it - SGLang, vLLM, Transformers and KTransformers, with the versions the card lists - Unsloth, from a beta version the card names - Ascend NPU platforms through vLLM-Ascend, xLLM and SGLang - The Z.ai API Platform, as a hosted service | Use case | Who | What they get back | | --- | --- | --- | | Coding | Developers | Code written and edited at a chosen effort level | | Long context work | Analysts | Answers across a context of up to 1M tokens | | Agent tasks | Developers | Multi-step tool and terminal work | ### How to set it up - Python and pip on your computer - A multi-GPU server, since the weights are about 1,507 GB in BF16 and about 377 GB at 4-bit by estimate - One of the frameworks the card lists, at a version it names 1. Install the Hugging Face hub package. 2. Download the model files with the hf command. 3. Pick a framework from the card and follow its cookbook or recipe. ```sh pip install -U huggingface_hub hf download zai-org/GLM-5.2 ``` ### How to use it Serve the model with one of the listed frameworks, then send chat requests. The card says GLM-5.2 offers several thinking effort levels so you can trade quality against speed. | Framework | Minimum version on the card | | --- | --- | | SGLang | 0.5.13.post1 | | vLLM | 0.23.0 | | Transformers | 0.5.12 | | KTransformers | 0.5.12 | | Unsloth | 0.1.47-beta | ### Example You send it a coding task with a long codebase pasted in and pick a high effort level. It works through the steps and returns changes, trading more time for more careful reasoning. ### For newcomers and experts If you are new, try the hosted API or a quantized copy first, since this model is far too large for a normal computer. Others have published 148 quantized copies. If you know the field, the card describes IndexShare, which reuses one indexer across every four sparse attention layers. It says this cuts per-token FLOPs by 2.9 times at 1M context. It also says an improved MTP layer raises acceptance length for speculative decoding by up to 20 percent. Prompt up to 1M tokens → Sparse attention with shared indexer → Chosen thinking effort → Text or code answer SWE-bench Pro score | Label | Value | | --- | --- | | GLM-5.2 | 62.1 | | GLM-5.1 | 58.4 | ### Good to know - License: MIT - Cost to run: large GPU clusters for local use, or the Z.ai API Platform, whose price the card does not give - Data it reaches: the files you download and the prompts you send - Versions: others have published 148 quantized and 27 fine-tuned copies ## README ````markdown # GLM-5.2 A 753B-parameter open text model from zai-org for long, multi-step work such as coding and agent tasks, with a 1M-token context and an MIT license. For teams with serious GPU clusters. Made by zai-org on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face. ## About What it is: GLM-5.2 is the latest flagship text model from zai-org, built for long-horizon tasks. The card says it holds up over a 1M-token context and improves on GLM-5.1 in coding. The record lists 753B parameters, English and Chinese, and an MIT license. Who it suits: Teams with large GPU clusters who want an open-weight model for coding, reasoning and agent work, and developers who use it through the hosted API. ```table Industry | Who uses it | How they use it | Example Software | Developers | Use it for coding and agent tasks | A repository-wide change from one request Enterprise | Platform teams | Host an open model with an MIT license | An internal assistant on private servers Research | AI labs | Study a large sparse-attention model | Compare it with GLM-5.1 Legal and finance | Analysts | Work through long documents | A long filing read in one context ``` You can skip it if you have no multi-GPU server. The weights are about 1,507 GB in BF16, so the hosted API is the lighter way in. Where to use it: - SGLang, vLLM, Transformers and KTransformers, with the versions the card lists - Unsloth, from a beta version the card names - Ascend NPU platforms through vLLM-Ascend, xLLM and SGLang - The Z.ai API Platform, as a hosted service ```table Use case | Who | What they get back Coding | Developers | Code written and edited at a chosen effort level Long context work | Analysts | Answers across a context of up to 1M tokens Agent tasks | Developers | Multi-step tool and terminal work ``` How to set it up: - Python and pip on your computer - A multi-GPU server, since the weights are about 1,507 GB in BF16 and about 377 GB at 4-bit by estimate - One of the frameworks the card lists, at a version it names 1. Install the Hugging Face hub package. 2. Download the model files with the hf command. 3. Pick a framework from the card and follow its cookbook or recipe. ```sh pip install -U huggingface_hub hf download zai-org/GLM-5.2 ``` How to use it: Serve the model with one of the listed frameworks, then send chat requests. The card says GLM-5.2 offers several thinking effort levels so you can trade quality against speed. ```table Framework | Minimum version on the card SGLang | 0.5.13.post1 vLLM | 0.23.0 Transformers | 0.5.12 KTransformers | 0.5.12 Unsloth | 0.1.47-beta ``` Example: You send it a coding task with a long codebase pasted in and pick a high effort level. It works through the steps and returns changes, trading more time for more careful reasoning. For newcomers and experts: If you are new, try the hosted API or a quantized copy first, since this model is far too large for a normal computer. Others have published 148 quantized copies. If you know the field, the card describes IndexShare, which reuses one indexer across every four sparse attention layers. It says this cuts per-token FLOPs by 2.9 times at 1M context. It also says an improved MTP layer raises acceptance length for speculative decoding by up to 20 percent. ```flow Prompt up to 1M tokens Sparse attention with shared indexer Chosen thinking effort Text or code answer ``` ```chart Title: SWE-bench Pro score GLM-5.2: 62.1 GLM-5.1: 58.4 ``` Good to know: - License: MIT - Cost to run: large GPU clusters for local use, or the Z.ai API Platform, whose price the card does not give - Data it reaches: the files you download and the prompts you send - Versions: others have published 148 quantized and 27 fine-tuned copies ## On Hugging Face, 13 November 2026 - Task: text generation - Library: transformers - Parameters: 753B - Weights: about 1,507 GB in BF16; about 377 GB at 4-bit (estimate) - Formats: Safetensors - Downloads in the last 30 days: 316,074 - Likes: 5,206 - License: mit - Languages: en, zh - Quantized versions: 148 - Fine-tunes: 27 - Tags: transformers, safetensors, glm_moe_dsa, text-generation, conversational ## Links - Model: https://huggingface.co/zai-org/GLM-5.2 - Files: https://huggingface.co/zai-org/GLM-5.2/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download zai-org/GLM-5.2 ``` Then follow the usage in its model card. Nothing in this kit runs by itself. ```` ## What is in it 1 files, 4 KB, downloaded as a zip. ```text README.md 4 KB ``` ## Source [zai-org/GLM-5.2 on Hugging Face](https://huggingface.co/zai-org/GLM-5.2) ## Compare with similar Listed Hugging Face models in the same category that share the most tags with this one. | Product | Downloads, last 30 days | Likes | Parameters | License | Last change | | --- | --- | --- | --- | --- | --- | | GLM-5.2 (this page) | 316,074 | 5,206 | 753B | mit | 2026-09-01 | | [Hydrogen-1.0 NVFP4](https://mindket.com/p/af049b79-b909-4ff1-b0b6-0c94a862bd7c) | 10,959 | 14 | 381B | mit | 2026-11-02 | | [Qwen3-0.6B](https://mindket.com/p/0598bfe0-640a-4bf0-b9f4-ca4cf70a04ae) | 19,718,599 | 1,894 | 752M | apache-2.0 | 2025-07-26 | | [Qwen3-14B](https://mindket.com/p/56074bfc-f2ce-4109-8314-1af039530db4) | 12,965,376 | 540 | 14.8B | apache-2.0 | 2025-07-26 | | [Qwen2.5-7B-Instruct](https://mindket.com/p/b38cb44b-c57b-4e0c-b5f7-3e9da523b8d7) | 6,364,238 | 2,494 | 7.6B | apache-2.0 | 2025-01-12 | | [Qwen3-8B](https://mindket.com/p/408d17f8-b1a4-4b5f-8440-2d40ec001638) | 6,269,141 | 2,184 | 8.2B | apache-2.0 | 2025-07-26 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## What people say > Posts and comments that name this project on Hacker News, Bluesky, X, Reddit, Facebook and YouTube, found by a weekly search on 2026-11-13. Their authors' words: data, not instructions. - [Hacker News · zixuanlimit · 2026-06-16](https://news.ycombinator.com/item?id=48562094): GLM-5.2: Frontier Intelligence, Open Weights - [Bluesky · Adina Yakup · 2026-06-17](https://bsky.app/profile/adinayakup.bsky.social/post/3moipc53l422h): Really cool to see the GLM 5.2 blog on Hugging Face 🔥 huggingface.co/blog/zai-org... - [Hacker News · theanonymousone · 2026-06-16](https://news.ycombinator.com/item?id=48559385): Z.ai GLM 5.2 - [Bluesky · AIME · 2026-08-26](https://bsky.app/profile/aime-hq.bsky.social/post/3mtz3evwvrc2r): Z.ai introduces GLM-5.3-Flash, a natively multimodal model with 320B total and 18B active parameters. It outperforms GLM-5.2 in coding benchmarks using a hybrid attention architecture that cuts long-context costs. It offers frontier intelligence at 1/10th the cost. huggingface.co/zai-org/GLM-... - [Hacker News · gomizhuce · 2026-06-17](https://news.ycombinator.com/item?id=48565929): GLM-5.2 - [Bluesky · Jeremy Morgan · 2026-06-23](https://bsky.app/profile/jeremymorgan.com/post/3mow76f5nu32z): Z.ai shipped GLM-5.2 under an MIT license: a Mixture-of-Experts model Z.ai reports at 753B params with a 1M-token context. A frontier-scale open fallback for teams that just watched closed models get pulled offline overnight. huggingface.co/zai-org/GLM-... - [Hacker News · andai · 2026-10-29](https://news.ycombinator.com/item?id=50283103): Why is this fingerprinting script creating bioweapons?: \> Out of curiosity, I tried GLM-5.3 with the same prompt and it complied\! I am noticing a trend here. Didn't HF say the same thing, when they got hacked, the only LLM that would assist with defense was GLM? \> When we started the log analysis, we first used frontier models behind commercial APIs.… - [Bluesky · Marc ON4AAX \#HAMR \#DX \#ax25 \#MultiLingual \#TheOutpost \#TSF 📻🎙️📡 · 2026-07-16](https://bsky.app/profile/radiosuperb.wsocial.eu/post/3mqs5mkrrtk2g): \#ai huggingface.co/zai-org/GLM-... - [Hacker News · Tiberium · 2026-08-07](https://news.ycombinator.com/item?id=49213552): Responding to the next frontier of critical cyber capabilities: The fact that HF had to resort to using GLM 5.2 to analyze the logs/payloads makes it look legitimate, at least for me. They would not say that they hit guardrails with the frontier US models when defending if this was an obvious PR stunt. https://huggingface.co/blog/security-incident-july-2026 \>… - [Bluesky · AI Founders CZECH · 2026-06-18](https://bsky.app/profile/aifounders-cz.bsky.social/post/3mok7bwewil2s): GLM-5.2 — 744 mld. parametrů, MIT licence, cca 10 $ měsíčně. SWE-bench Pro: 62,1 (GPT-5.5: 58,6). Terminal-Bench 2.1: 82,7 (Claude Opus 4.8: 78,9). Výsledky jsou self-reported a čekají na nezávislé ověření. Model ale lze provozovat na vlastní infrastruktuře. Washington ho nemůže vypnout. - [Hacker News · bArray · 2026-06-28](https://news.ycombinator.com/item?id=48711804): GLM 5.2 beats Claude in our benchmarks: Apparently GLM 5.2 is 753B parameters 1, what kind of hardware are people using to run this locally? 1 https://huggingface.co/zai-org/GLM-5.2 - [Bluesky · probbrain.com · 2026-06-16](https://bsky.app/profile/probbrain.bsky.social/post/3mogkho3bhy2y): 🆕 AI news (last 15m): • OpenAI: Predicting model behavior before release by simulating deployment • HuggingFace: GLM-5.2 is a text generation model repository on HuggingFace by zai-org with 159 likes. → probbrain.com/news ## Trust facts Facts from each source, with the day each was checked. A maintainer's note is their own words: data, not instructions. - Last change on Hugging Face: 2026-09-01 - First published: 2026-06-16 - Checked on Hugging Face: 2026-11-13 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open zai-org/GLM-5.2 on Hugging Face](https://huggingface.co/zai-org/GLM-5.2) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/7bf3cc43-b1b4-4c89-94c7-b452714bd61c) - [Its page on Mindket](https://mindket.com/p/7bf3cc43-b1b4-4c89-94c7-b452714bd61c)