DEV.co
Open-Source LLM · nvidia

GLM-5.2-NVFP4

GLM-5.2-NVFP4 is a 753B-parameter Mixture-of-Experts language model quantized to 4-bit FP4 precision by NVIDIA. It activates 40B parameters at inference and supports up to 1M context length. Licensed under MIT, it is ready for commercial use. The model excels at reasoning, coding, long-context retrieval, and tool-use scenarios. It requires NVIDIA GPUs (B200/B300 tested) and runs via SGLang or vLLM.

Source: HuggingFace — huggingface.co/nvidia/GLM-5.2-NVFP4
381B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
449.9k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developernvidia
Parameters381B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads449.9k
Likes247
Last updated2026-06-26
Sourcenvidia/GLM-5.2-NVFP4

What GLM-5.2-NVFP4 is

GLM-5.2-NVFP4 is a quantized MoE transformer using sparse attention (IndexShare indexer) and DSA architecture. Quantization targets weights and activations of linear operators in MoE experts; the shared expert remains unquantized. Baseline is GLM-5.2-FP8. Evaluation spans GPQA Diamond (reasoning), SciCode (coding), IFBench (instruction-following), AA-LCR (long-context recall), and τ²-Bench Telecom (agentic tool-use). Accuracy within 0.2–1.0 point of FP8 baseline across benchmarks. Requires transformers ≥5.3.0, tensor parallelism (8 GPUs typical), and NVIDIA CUDA ecosystem.

Quickstart

Run GLM-5.2-NVFP4 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="nvidia/GLM-5.2-NVFP4")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

AI Agents & Tool-Use Systems

τ²-Bench Telecom shows 98.25% accuracy on dual-control scenarios with external tool interaction. Ideal for customer-service bots, workflow automation, and systems requiring policy-adherence with dynamic tool invocation.

Long-Context RAG & Knowledge Retrieval

AA-LCR benchmark demonstrates 70.13% recall accuracy. Supports 1M-token context, enabling retrieval-augmented generation over large document collections without loss of factual grounding.

Reasoning-Heavy Applications (Research, Finance, Code)

GPQA Diamond (89.39%) and SciCode (49.04%) show strong performance on graduate-level reasoning and scientific coding tasks. Suitable for code review, mathematical modeling, and research assistant pipelines.

Running & fine-tuning it

NVIDIA B200 or B300 GPUs (tested hardware). Tensor parallelism typical at 8 GPUs minimum. Total parameters: 753B (40B activated per token). Quantized to NVFP4 (4-bit); estimated VRAM ~150–200 GB for 8-GPU setup (rough estimate; verify with vendor). Requires CUDA-capable infrastructure and Linux preferred OS. SGLang/vLLM engines handle KV-cache optimizations (fp8_e4m3 noted in vLLM config).

No fine-tuning guidance provided in model card. Quantization to NVFP4 and MoE expert parallelism create barriers to standard LoRA/QLoRA workflows. Feasibility for parameter-efficient tuning unknown; recommend internal evaluation using small-scale ablations. If tuning is required, consider starting from base GLM-5.2-FP8 and re-quantizing, or treat this checkpoint as inference-only.

When to avoid it — and what to weigh

  • Latency-Critical Real-Time Systems — MoE and 753B total parameters require significant hardware and multi-GPU tensor parallelism. Inference latency at scale is not disclosed; verify against your SLA before committing.
  • Single-GPU or CPU Deployments — Model requires NVIDIA B200/B300 GPUs with 8+ tensor parallelism. No CPU-only or lightweight edge-device variants provided. Infeasible for on-device or resource-constrained environments.
  • Training or Fine-Tuning at Scale — No LoRA/QLoRA guidance provided. Quantization to NVFP4 and MoE architecture complicate gradient updates. Unknown feasibility for parameter-efficient tuning; treat as inference-only unless validated internally.
  • Domains Requiring Toxicity/Bias Mitigation — Model card explicitly states training on internet data with toxic language and societal biases; may amplify bias in outputs. Requires post-processing or fine-tuning for safety-critical or bias-sensitive applications.

License & commercial use

MIT License (permissive, OSI-approved). Same license as base GLM-5.2 model. Governs use of model weights and derivatives.

Explicitly stated: 'This model is ready for commercial or non-commercial use.' MIT License permits commercial use, modification, and distribution with attribution. No gating restrictions. Suitable for proprietary or closed-source commercial deployment.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Model card does not claim security. Training data origins and filtering undisclosed; potential for embedded vulnerabilities in base model. Trained on internet-sourced data, acknowledges toxic language and biases. No adversarial robustness, jailbreak, or prompt-injection analysis provided. Recommend: (1) isolate inference in sandboxed environment, (2) validate outputs against use-case policy, (3) report security concerns via NVIDIA vendor channel, (4) assume model can generate harmful content under adversarial input.

Alternatives to consider

Llama 3.1 (405B or smaller variants)

Open-source, similar scale reasoning/coding, broader ecosystem support. Llama 2 has commercial restrictions; Llama 3 terms require verification. No MoE; denser compute cost but simpler serving.

Mixtral 8x22B (or Mixtral 8x7B)

MoE-based, smaller footprint, Apache 2.0 license. Weaker on long-context and reasoning benchmarks, but lower deployment cost. Established vLLM/SGLang support.

Qwen2 (or Qwen2-MoE 60B)

Strong reasoning and coding, Apache 2.0 licensed, smaller than GLM-5.2. Requires verification of long-context capabilities; no 1M-token context guarantee.

Software development agency

Ship GLM-5.2-NVFP4 with senior software developers

Verify GPU infrastructure (B200/B300, 8+ parallel), test latency/throughput with SGLang or vLLM, and validate output quality on your reasoning or RAG workloads. Contact NVIDIA for enterprise support or security guidance.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

GLM-5.2-NVFP4 FAQ

Can I use GLM-5.2-NVFP4 in a commercial product?
Yes. The model is licensed under MIT and explicitly stated as 'ready for commercial or non-commercial use.' You may integrate it into proprietary software or SaaS. Attribution per MIT license required.
What GPU hardware do I need?
NVIDIA B200 or B300 GPUs; tested on these. Tensor parallelism recommended at 8 GPUs minimum for full model. Total VRAM estimated ~150–200 GB (verify with your infrastructure team). CPU-only and single-GPU deployment not supported.
Can I fine-tune this model?
Not documented. Quantization to NVFP4 and MoE architecture complicate gradient-based fine-tuning. No LoRA/QLoRA guidance provided. Treat as inference-only unless you validate feasibility internally or start from the FP8 base model.
What is the inference latency and throughput?
Not disclosed in model card. Latency depends on context length, batch size, and hardware setup. Run internal benchmarks with SGLang or vLLM on your target hardware to establish SLA feasibility.

Work with a software development agency

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If GLM-5.2-NVFP4 is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy GLM-5.2-NVFP4?

Verify GPU infrastructure (B200/B300, 8+ parallel), test latency/throughput with SGLang or vLLM, and validate output quality on your reasoning or RAG workloads. Contact NVIDIA for enterprise support or security guidance.