DeepSeek V4.1 Flash

Model ID: vllm/release/deepseek-v41-flash DeepSeek Experimental

EU AI Act

Attribute	Value
Classification	GPAI (General-Purpose AI Model)
Risk class	Limited Risk (Article 50)
Transparency obligation	Yes
noris role	Distributor

As a GPAI model, this model falls within the scope of the EU AI Act transparency obligations. noris acts as a distributor and makes the model developer’s information available pursuant to Article 53(2) and (3).

Technical Specifications

Property	Value
Model name	DeepSeek V4.1 Flash
Provider	DeepSeek
Model ID	vllm/release/deepseek-v41-flash
Parameters	552B backbone (MoE), 8B active during prefill, 16B during decode, plus 196B Engram
Architecture	Causal encoder-decoder MoE with Compressed Sparse Attention 2 (CSA2)
Context length	1,048,576 tokens
Modalities	Text + image (multimodal)
Checkpoint	NVFP4, quantized with NVIDIA Model Optimizer
License	MIT

Recommended Use Cases

• AI agents and long-running tasks (Terminal-Bench 2.1: 82.2%)
• Chatbots and assistants
• RAG systems
• Very-long-document analysis (1M context)
• Image understanding (scientific charts, MMMU-Pro: 73.7%)

Known Limitations

Caution

• Very large model (high latency): due to its size, inference is slower than with smaller models.
• Experimental tier (2-week deprecation notice): this model may be deprecated with only two weeks advance notice.
• Internet-derived training data: the model may produce biased or toxic answers, especially when prompted accordingly.

Risk Assessment (EU AI Act)

Risk Factor	Category	Level
Model size	Deepfake potential	Medium
Multimodal	Image input	Medium
Open source (MIT)	Misuse potential	Low
Experimental tier	Deprecation notice	Flagged
Environmental impact	Model size (552B backbone)	Medium-High

Training Data & Sustainability

Training data: not fully publicly documented, see the base model card. The NVFP4 checkpoint was calibrated with cnn_dailymail and the Nemotron-Post-Training-Dataset-v2.

Sustainability:

Metric	Value
CO₂	High (552B backbone)
Optimization	only 8B (prefill) resp. 16B (decode) active parameters despite the large expert pool

Original Sources

• Base model: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
• NVFP4 checkpoint: https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-NVFP4
• License: MIT License

Model card created on October 7, 2026. This model card serves EU AI Act transparency purposes (Regulation (EU) 2024/1689). The information is based on publicly available data from the model provider.