GLM-5.2-NVFP4 by NVIDIA: Open-Weight Model

Runs On

What it takes to serve GLM-5.2-NVFP4 (381B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

Precision	Weights	Memory needed	Cheapest setup	Per hour	Also fits
16-bit	762.0 GB	914.4 GB	4x MI325X (256 GB) Vultr	$8.00	5x MI300X $9.25 · 4x MI355X $10.36
8-bit	381.0 GB	457.2 GB	2x MI325X (256 GB) Vultr	$4.00	2x MI355X $5.18 · 3x MI300X $5.55
4-bit	190.5 GB	228.6 GB	1x MI325X (256 GB) Vultr	$2.00	1x MI355X $2.59 · 2x MI300X $3.70

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 10, 2026.

GLM-5.2-NVFP4 on every accelerator the SAVRN Index prices, at every precision

SAVRN's Notes on GLM-5.2-NVFP4

381B parameters, 256 experts with 8 active per token, a 1,048,576-token context: this is a frontier-scale mixture-of-experts, and the Runs On table prices it accordingly. At 16-bit you need 914.4 GB — four MI325X cards at $8.00 an hour. At 8-bit, two MI325X at $4.00. At 4-bit, one MI325X at $2.00. The 464.8 GB download arrives as 47 shards of roughly 10 GB each, so budget disk and staging time before you budget GPU.

The license is MIT, unusually clean for this class. The name says NVFP4 and the library column says Model Optimizer — this is NVIDIA's quantized repack of zai-org/GLM-5.2 — while our generic precision table reports the standard 16/8/4-bit lines above, not a measured NVFP4 figure. Treat the 4-bit row as the planning number and the publisher's own calibration as the ceiling. Neighbors on the hub at this scale: K2-Horizon-375B and Qwen3-Coder-480B-FP8.

SAVRN Research

Model Card

By NVIDIA, published under mit, revision 53e0691e2189.

Model Overview

Description:

The NVIDIA GLM-5.2 NVFP4 model is the quantized version of ZAI’s GLM-5.2 model, which is an auto-regressive language model that uses an optimized transformer architecture. GLM-5.2 is a Mixture-of-Experts (MoE) model for reasoning and coding that uses sparse attention (with an IndexShare indexer) to support a long context. For more information, please check here. The NVIDIA GLM-5.2 NVFP4 model is quantized with Model Optimizer.

This model is ready for commercial or non-commercial use.

License/Terms of Use:

GOVERNING TERMS: Use of the model is governed by the MIT License, same as the base model.

Deployment Geography:

Global

Use Case:

Developers looking to take off-the-shelf, pre-quantized models for deployment in AI Agent systems, chatbots, RAG systems, and other AI-powered applications.

Release Date:

Hugging Face 06/25/2026 via https://huggingface.co/nvidia/GLM-5.2-NVFP4

References

Nvidia Model Optimizer: https://github.com/NVIDIA/Model-Optimizer

Model Architecture:

Architecture Type: Transformers
Network Architecture: GLM-5.2 (GlmMoeDsaForCausalLM)
Number of Model Parameters: 753B in total and 40B activated

Input:

Read the full model card (1,412 words)

Configuration

Architecture: GlmMoeDsaForCausalLM
Context length (tokens): 1,048,576
Layers: 78
Hidden size: 6,144
Feed-forward size: 12,288
Attention heads: 64
Key/value heads: 64
Head dimension: 192
Vocabulary size: 154,880
Routed experts: 256
Experts: 256
Experts active per token: 8
Model type: glm_moe_dsa
Quantization: modelopt

Identity and Version

Repository: nvidia/GLM-5.2-NVFP4
Publisher: NVIDIA
Task: Text generation
Modality: Text
Library: Model Optimizer
Parameters: 381B parameters
Languages: Not stated by the source
Revision: 53e0691e21895a3863a606dfd12910c69eba94ab
First published: 2026-06-22
Last updated: 2026-08-31

Files and Weights

56 files, 464.9 GB in total. The weights are 47 files totalling 464.8 GB in safetensors.

Weights47 files · 464.8 GB

Configuration4 files · 22.2 MB

Tokenizer2 files · 20.2 MB

Documentation1 file · 12.4 KB

Other1 file · 5.1 KB

Repository1 file · 1.6 KB

Every file

File	Type	Size	SHA-256
model-00001-of-00047.safetensors	Weights	10.0 GB	7355ca8065da
model-00002-of-00047.safetensors	Weights	10.0 GB	5a93bf6840c9
model-00003-of-00047.safetensors	Weights	10.0 GB	240485288e04
model-00004-of-00047.safetensors	Weights	10.0 GB	790fe046fb8b
model-00005-of-00047.safetensors	Weights	10.0 GB	9c6a3381b16f
model-00006-of-00047.safetensors	Weights	10.0 GB	902e254158f0
model-00007-of-00047.safetensors	Weights	10.0 GB	40299690b696
model-00008-of-00047.safetensors	Weights	10.0 GB	139f1c107077
model-00009-of-00047.safetensors	Weights	10.0 GB	e1db5be0dac7
model-00010-of-00047.safetensors	Weights	10.0 GB	617a2ca6fd37
model-00011-of-00047.safetensors	Weights	10.0 GB	5099883c548f
model-00012-of-00047.safetensors	Weights	10.0 GB	df6a0c6ab9fa
model-00013-of-00047.safetensors	Weights	10.0 GB	2bfd1392eda1
model-00014-of-00047.safetensors	Weights	10.0 GB	6320d889ab91
model-00015-of-00047.safetensors	Weights	10.0 GB	499d121db3e7
model-00016-of-00047.safetensors	Weights	10.0 GB	b0bfeb109e4a
model-00017-of-00047.safetensors	Weights	10.0 GB	ca88817f6f02
model-00018-of-00047.safetensors	Weights	10.0 GB	4d32422c750f
model-00019-of-00047.safetensors	Weights	10.0 GB	74ecfce9cddd
model-00020-of-00047.safetensors	Weights	10.0 GB	23a23304160c
model-00021-of-00047.safetensors	Weights	10.0 GB	db14cdfbe3e5
model-00022-of-00047.safetensors	Weights	10.0 GB	bf5f1c293e95
model-00023-of-00047.safetensors	Weights	10.0 GB	2fb70768ccd1
model-00024-of-00047.safetensors	Weights	10.0 GB	ccf3b2e4ca1e
model-00025-of-00047.safetensors	Weights	10.0 GB	67f233905c40
model-00026-of-00047.safetensors	Weights	10.0 GB	daf23e3ccbf4
model-00027-of-00047.safetensors	Weights	10.0 GB	1b4c7404684d
model-00028-of-00047.safetensors	Weights	10.0 GB	a895f3e5a815
model-00029-of-00047.safetensors	Weights	10.0 GB	5601bf446f5f
model-00030-of-00047.safetensors	Weights	10.0 GB	355b7b4665a8
model-00031-of-00047.safetensors	Weights	10.0 GB	7bc1ca71bffd
model-00032-of-00047.safetensors	Weights	10.0 GB	d11bd3fe6f4c
model-00033-of-00047.safetensors	Weights	10.0 GB	e39c9ff0d42c
model-00034-of-00047.safetensors	Weights	10.0 GB	a81b529612c9
model-00035-of-00047.safetensors	Weights	10.0 GB	70d19f0b6180
model-00036-of-00047.safetensors	Weights	10.0 GB	cab4d531d014
model-00037-of-00047.safetensors	Weights	10.0 GB	dc213d3550fc
model-00038-of-00047.safetensors	Weights	10.0 GB	50c4a5522bd9
model-00039-of-00047.safetensors	Weights	10.0 GB	ca17fd221731
model-00040-of-00047.safetensors	Weights	10.0 GB	103a33ce23be
model-00041-of-00047.safetensors	Weights	10.0 GB	4d398e69fdd5
model-00042-of-00047.safetensors	Weights	9.9 GB	573afa976e2f
model-00043-of-00047.safetensors	Weights	10.0 GB	a459ccf072fc
model-00044-of-00047.safetensors	Weights	10.0 GB	5ef1a45b4d69
model-00045-of-00047.safetensors	Weights	10.0 GB	0fa60cd2ffdd
model-00046-of-00047.safetensors	Weights	10.0 GB	6edd678c7c89
model-00047-of-00047.safetensors	Weights	5.1 GB	ca2dfdb8f16f
config.json	Configuration	15.5 KB	—
generation_config.json	Configuration	215 B	—
hf_quant_config.json	Configuration	7.4 KB	—
model.safetensors.index.json	Configuration	22.2 MB	2aa8397b501d
README.md	Documentation	12.4 KB	—
chat_template.jinja	Other	5.1 KB	—
.gitattributes	Repository	1.6 KB	—
tokenizer.json	Tokenizer	20.2 MB	19e773648cb4
tokenizer_config.json	Tokenizer	790 B	—

License and Download

License: mit
Access: Open weights, no gate
Download size: 464.8 GB

Download from NVIDIA

Released by NVIDIA through its official repository on Hugging Face. Read the license.

Built From

Derived from zai-org/GLM-5.2
Quantized from zai-org/GLM-5.2

Memory Requirements

Precision	Weights in memory
As published	464.8 GB
16-bit	762.0 GB
8-bit	381.0 GB
4-bit	190.5 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Compare GLM-5.2-NVFP4

DeepSeek-V4-Flash-0731 vs GLM-5.2-NVFP4

Questions About GLM-5.2-NVFP4

How much GPU memory does GLM-5.2-NVFP4 need?

About 914.4 GB at 16-bit and 228.6 GB at 4-bit: the weights (381B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run GLM-5.2-NVFP4 on?

At 16-bit, 4x MI325X from $8.00 an hour; at 4-bit, 1x MI325X from $2.00 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use GLM-5.2-NVFP4 commercially?

Yes. GLM-5.2-NVFP4 is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

What is GLM-5.2-NVFP4's context length?

1,048,576 tokens, from the maximum position embeddings in its published configuration.

Similar Models

Model · Text generation

K2-Horizon-375B-A23B

Institute of Foundation Models

K2-Horizon-375B-A23B is the flagship of the K2-Horizon family: a sparse Mixture-of-Experts model that stores 375B parameters and runs 23B per token, with a 512K context window. We have released the final checkpoint; intermediate checkpoints, along with the data and the training code, will be released. - Frontier-class agentic performance. On agentic tool use, terminal, and long-horizon workflow benchmarks it matches or beats open-weight MoE models up to 2.6× its size and is competitive with closed frontier models (see Benchmark Results). - 512K context. Native 524,288-token context from the midtraining stages onward. - Intermediate checkpoints. Intermediate checkpoints will be released so…

Open weights apache-2.0 379.2B parameters 524,288 tokens transformers

View model

Model · Text generation

Darwin-397B-JGOS

FINAL_Bench

Darwin-397B-JGOS is the largest and highest-scoring member of the Darwin family. Built on Qwen 3.5 397B as the base, it transplants the FFN (expert) strengths of multiple high-performance models through the Darwin V9 platform, producing a 397B-parameter Mixture-of-Experts model with ~17B active parameters per token. It reaches 90.9 % on GPQA Diamond with pure greedy decoding (single sample) — surpassing Darwin-28B-REASON (89.39 %, achieved with the Darwin-DELPHI test-time engine) without using any test-time engine at all. This is the highest GPQA Diamond score in the Darwin family to date. Darwin is VIDRAFT's measuring-result-driven reasoning model family — approximately 20 official models…

Open weights apache-2.0 403.4B parameters 262,144 tokens transformers

View model

Model · Text generation

Darwin-397B-ZTC

FINAL_Bench

the footprint, GPQA Diamond 93.43 %. And this model stops itself before it acts on an answer it is about to get wrong. Darwin is VIDRAFT's measurement-driven reasoning model family — roughly 20 official models, 400+ community derivatives, and a standing place among the top open models on GPQA. A large MoE model is made of hundreds of experts. Darwin V9 selects the experts that perform best across several high-performing models, transplants them onto a base backbone, and fuses them with trust-weighted evolutionary merging. Nothing is trained from scratch — proven capability is grafted on. That is why the same method holds across every model size. - Darwin V9 — evolutionary FFN/expert…

Open weights apache-2.0 403.6B parameters 262,144 tokens transformers

View model

Model · Text generation

DeepSeek-V4-Flash-0731

DeepSeek

DeepSeek-V4-Flash-0731 is the official release of DeepSeek-V4-Flash, superseding the preview version, with substantially enhanced agentic capabilities. It has the same model structure as DeepSeek-V4-Flash-DSpark, i.e. it comes with a speculative decoding module attached. DeepSeek-V4-Flash-0731 outperforms DeepSeek-V4-Pro (Preview) on benchmarks listed below despite its far smaller activated parameter count, and is broadly competitive with the strongest proprietary models available. 1. For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework, using the max reasoning effort…

Open weights mit 304.2B parameters 1,048,576 tokens transformers

View model

Model · Text generation

DeepSeek-V4-Flash

DeepSeek

We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens. DeepSeek-V4 series incorporate several key upgrades in architecture and optimization: 1. Hybrid Attention Architecture: We design a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to dramatically improve long-context efficiency. In the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache…

Open weights mit 290.9B parameters 1,048,576 tokens transformers

View model

Model · Text generation

Qwen3-Coder-480B-A35B-Instruct-FP8

Qwen

Today, we're announcing Qwen3-Coder, our most agentic code model to date. Qwen3-Coder is available in multiple sizes, but we're excited to introduce its most powerful variant first: Qwen3-Coder-480B-A35B-Instruct. featuring the following key enhancements: - Significant Performance among open models on Agentic Coding, Agentic Browser-Use, and other foundational coding tasks, achieving results comparable to Claude Sonnet. - Long-context Capabilities with native support for 256K tokens, extendable up to 1M tokens using Yarn, optimized for repository-scale understanding. - Agentic Coding supporting for most platform such as Qwen Code, CLINE, featuring a specially designed function call format.…

Open weights apache-2.0 480.2B parameters 262,144 tokens transformers

View model