Modelpublic

bg-digitalservices/Apertus-70B-Instruct-2509-NVFP4

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face

2likes27downloads

Model Card

Apertus-70B-Instruct-2509-NVFP4

NVFP4 quantization of swiss-ai/Apertus-70B-Instruct-2509 — part of the Swiss AI Apertus model family. 70B dense transformer supporting 1,811 languages with 65K context.

W4A4 — both weights and activations in FP4. Maximum speed on Blackwell GPUs. See also [Apertus-70B-Instruct-2509-NVFP4A16](https://huggingface.co/bg-digitalservices/Apertus-70B-Instruct-2509-NVFP4A16) for the weight-only W4A16 variant.

Key Specs

	Original (BF16)	NVFP4 (this)
Size on disk	~140 GB	~35 GB
Compression	—	~3.0x
Parameters	70B	70B
Architecture	Dense transformer, xIELU activation	same
Context window	65,536 tokens	65,536 tokens
Languages	1,811	1,811

Serving with vLLM

bash

vllm serve bg-digitalservices/Apertus-70B-Instruct-2509-NVFP4 \
  --quantization modelopt \
  --dtype auto \
  --kv-cache-dtype fp8 \
  --gpu-memory-utilization 0.85 \
  --max-model-len 65536 \
  --trust-remote-code

DGX Spark

bash

VLLM_NVFP4_GEMM_BACKEND=marlin vllm serve bg-digitalservices/Apertus-70B-Instruct-2509-NVFP4 \
  --quantization modelopt \
  --dtype auto \
  --kv-cache-dtype fp8 \
  --max-model-len 65536 \
  --trust-remote-code

Testing

This is an instruct model with tool use support — use the chat completions endpoint.

Quantization Details

—Method: NVIDIA Model Optimizer (modelopt) v0.43
—Format: NVFP4 — E2M1 weights with per-group FP8 scales (group size 16)
—Calibration: 4096 samples from CNN/DailyMail, batch size 32, seq_len 1024
—Hardware: NVIDIA H200 GPU
—Quantization script: included as quantize.py

About Apertus

Apertus is built by Swiss AI — a fully open, privacy-first model family trained on 4,096 GH200 GPUs. Key features:

—1,811 native languages
—Novel xIELU activation + AdEMAMix optimizer
—EU AI Act compliant, respects opt-out consent
—Full training transparency (weights, data, scripts all public)

License

Apache 2.0 — inherited from the base model.

Citation

If you use this model, please cite the original Apertus work:

bibtex

@misc{swisstransformer2025apertus,
  title   = {Apertus},
  author  = {Swiss Transformer},
  year    = {2025},
  url     = {https://huggingface.co/swiss-ai}
}

Credits

Quantized by Mario Iseli on an NVIDIA H200. Built and validated with AI-engineering assistance from Anthropic.

📬 mario@marioiseli.com ☕ Buy me a coffee if this makes your inference go brrrrrr! 🚀