How Nvidia’s NVFP4 Cuts GLM‑5.2 Deployment Cost by Half

Semgrep security benchmark

Semgrep evaluated several code‑oriented LLMs on an IDOR detection benchmark using a pure‑prompt (no harness) setup. The ranking by F1 score was:

Semgrep Multimodal (GPT‑5.5) – full harness – 61%

Semgrep Multimodal (Opus 4.8) – full harness – 53%

GLM‑5.2 – pure prompt – 39%

Claude Code (Opus 4.6) – Claude Code SDK – 37%

Claude Code (Opus 4.8/4.7) – Claude Code SDK – 28%

MiniMax M3 – pure prompt – 23%

Kimi K2.7 Code – pure prompt – 22%

GPT‑5.5 – Codex – 20%

GLM‑5.2 discovered a vulnerability for $0.17, roughly one‑sixth the cost of the top closed‑source models.

Model architecture and core capabilities

GLM‑5.2 is Zhipu’s flagship open‑source model. It uses a 753 B parameter Mixture‑of‑Experts (MoE) architecture with 40 B active tokens per step. The proprietary IndexShare sparse‑attention mechanism enables a true 1 M‑token context window while reducing FLOPs per token by 2.9×.

Code benchmarks:

SWE‑bench Pro – 62.1 (Claude Opus 4.8: 69.2, GPT‑5.5: 58.6)

Terminal Bench – 81.0 (Claude: 85)

FrontierSWE – 74.4 (GPT‑5.5: 72.6)

Reasoning benchmarks:

AIME 2026 math – 99.2

GPQA Diamond scientific reasoning – 91.2

Even with FP8 precision, the 753 B MoE model requires at least eight high‑end GPUs for deployment.

NVIDIA NVFP4 quantization

On June 25, NVIDIA released nvidia/GLM-5.2-NVFP4 on Hugging Face. The model was quantized with NVIDIA Model Optimizer (v0.46.0), compressing weights and activations from FP8 to FP4. Only the linear operators inside expert‑layer Transformer blocks are quantized; shared‑expert layers retain original precision, minimizing accuracy loss.

Benchmark comparison (higher is better):

GPQA Diamond – FP8: 89.52, NVFP4: 89.39

SciCode – FP8: 49.85, NVFP4: 49.04

IFBench – FP8: 74.95, NVFP4: 75.81

AA‑LCR – FP8: 69.38, NVFP4: 70.13

τ²‑Bench Telecom – FP8: 97.90, NVFP4: 98.25

The NVFP4 version shows a negligible 0.13 drop on GPQA Diamond (within measurement error) and improves IFBench and τ²‑Bench Telecom scores, demonstrating that halving precision can retain or even increase performance on certain metrics.

Deployment options

Option 1: SGLang (officially recommended)

pip install -U "transformers>=5.3.0" && \
python3 -m sglang.launch_server \
    --model nvidia/GLM-5.2-NVFP4 \
    --tensor-parallel-size 8 \
    --quantization modelopt_fp4 \
    --tool-call-parser glm47 \
    --reasoning-parser glm45 \
    --trust-remote-code \
    --chunked-prefill-size 16384 \
    --mem-fraction-static 0.80

Option 2: vLLM

vllm serve nvidia/GLM-5.2-NVFP4 \
    --tensor-parallel-size 8 \
    --enable-expert-parallel \
    --trust-remote-code \
    --reasoning-parser glm45 \
    --tool-call-parser glm47 \
    --enable-auto-tool-choice \
    --kv-cache-dtype fp8_e4m3 \
    --host 0.0.0.0 --port 8000

Key flags: --enable-expert-parallel activates MoE expert parallelism. --kv-cache-dtype fp8_e4m3 compresses the KV cache to FP8. --tool-call-parser glm47 and --reasoning-parser glm45 match GLM‑5.2’s tool‑call and reasoning formats.

Official testing used Blackwell B200/B300 GPUs; these provide native FP4 compute for maximum throughput. Hopper‑based GPUs can run the model but do not benefit from FP4 acceleration.

NVIDIA Model Optimizer capabilities

The optimizer supports a range of techniques:

Post‑training quantization (PTQ) – 2‑4× model size reduction.

Quantization‑aware training (QAT) – restores accuracy after quantization.

Pruning – removes unimportant weights.

Distillation – trains smaller models with guidance from larger ones.

Speculative decoding – predicts draft tokens to lower latency.

Sparsification – stores only non‑zero parameters.

Installation: pip install -U nvidia-modelopt[all] Quantized models can be deployed to SGLang, vLLM, TensorRT‑LLM and other mainstream inference engines without additional conversion steps.

Conclusion

GLM‑5.2 is currently the strongest open‑source LLM for code and reasoning tasks. Semgrep’s benchmark confirms its practical superiority over Claude Code at roughly one‑sixth the cost per discovered vulnerability. NVIDIA’s NVFP4 quantization halves the memory footprint while keeping accuracy virtually unchanged and even improving some metrics. For teams with access to Blackwell‑based GPUs, the NVFP4‑quantized GLM‑5.2 provides the best price‑performance ratio for local AI agents, RAG systems, or workloads requiring 1 M‑token context windows.