DeepSeek-V4.1-Flash-NVFP4 — Serverless API on Token Factory by packet.ai
Run DeepSeek-V4.1-Flash (NVFP4) through a serverless, OpenAI-compatible API with pay-per-token billing. Token Factory by packet.ai serves open-weight models behind an OpenAI-compatible endpoint, so you can switch your existing OpenAI SDK code by changing only the base URL and API key.
This card describes the hosted API for nvidia/DeepSeek-V4.1-Flash-NVFP4, NVIDIA's 4-bit (NVFP4) quantization of deepseek-ai/DeepSeek-V4.1-Flash — a mixture-of-experts reasoning model with a 1M-token context window and image input.
Why use DeepSeek-V4.1-Flash-NVFP4 on Token Factory
- OpenAI-compatible endpoint — works with existing OpenAI SDKs.
- Streaming, JSON mode and tool calling supported on chat completions.
- Pay per token — input and output tokens are metered separately; nothing is charged while idle.
- Free for one month at launch (till Dec 16th), then billed per 1M tokens. See current pricing.
- Scale to dedicated GPUs — move the model to a dedicated GPU running vLLM while keeping the same endpoint shape.
Good for
- Reasoning (1M-token context, image input)
- Per NVIDIA's model card: AI agent systems, chatbots, RAG systems and other AI-powered applications
Model overview
| Property | Value |
|---|---|
| Model | DeepSeek-V4.1-Flash-NVFP4 |
| Base model | deepseek-ai/DeepSeek-V4.1-Flash |
| Quantized by | NVIDIA, using NVIDIA Model Optimizer |
| Architecture | Mixture-of-experts transformer with Compressed Sparse Attention 2 |
| Parameters | 552B backbone (8B activated in prefill, 16B in decode) + 196B Engram conditional memory |
| Context window | Up to 1M tokens |
| Input | Text, image |
| Output | Text |
| Quantization | Routed MoE experts in NVFP4 (W4A4, group size 16); Engram tables in MXFP4; attention, shared experts and vision kept at source precision |
| Target hardware | NVIDIA Blackwell |
| License | MIT |
Accuracy
Reported by NVIDIA (vLLM), comparing the NVFP4 checkpoint with the MXFP4 source model. See the original model card for evaluation settings.
| Precision | GPQA Diamond | AA-LCR | SciCode | IFBench | MMMU-Pro | Terminal-Bench 2.1 |
|---|---|---|---|---|---|---|
| MXFP4 (source) | 91.035 | 78.563 | 54.401 | 76.667 | 74.046 | 81.60 |
| NVFP4 + MXFP4 Engram | 90.783 | 79.562 | 53.994 | 77.867 | 74.566 | 81.88 |
Quickstart
1. Get an API key
Sign up at dash.packet.ai and create an API key.
2. Call the API (Python, OpenAI SDK)
from openai import OpenAI
client = OpenAI(
api_key="PACKET_API_KEY",
base_url="https://api.packet.ai/v1",
)
response = client.chat.completions.create(
model="nvidia-deepseek-v4-1-flash-nvfp4",
messages=[{"role": "user", "content": "Implement Hello World"}],
)
print(response.choices[0].message.content)
Replace PACKET_API_KEY with your API key. Full API documentation: dash.packet.ai/docs
Other models on Token Factory
| Model | Highlights |
|---|---|
MiniMaxAI/MiniMax-M3.1-Flash |
Agentic model for coding and computer use, 1M-token context, text/image/video input |
Qwen/Qwen2.5-1.5B-Instruct |
Small, fast model for classification and extraction, 32K context, 29+ languages |
See the full, up-to-date list at packet.ai/token-factory.
Run it yourself
Prefer to self-host? Use the original checkpoints:
- NVFP4 (NVIDIA Blackwell):
nvidia/DeepSeek-V4.1-Flash-NVFP4— see its card for vLLM and SGLang instructions - Original model:
deepseek-ai/DeepSeek-V4.1-Flash
About packet.ai
packet.ai is a hosted·ai project operated by HOSTED AI Inc., 2880 Zanker Rd, Suite 203, San Jose, CA 95134. Alongside Token Factory, packet.ai offers GPUs (B200, RTX PRO 6000, A100, RTX 4090), dynamic and dedicated GPUs, and bare metal/clusters, billed by the minute.
- Website: packet.ai
- Token Factory: packet.ai/token-factory
- Docs: dash.packet.ai/docs
- Talk to an engineer: packet.ai/contact
- X: @packet_ai · LinkedIn: packet-dot-ai · GitHub: hosted-ai
Limitations
As noted by NVIDIA, the base model was trained on web data that may contain toxic language and societal biases, and it can produce inaccurate or irrelevant output. Evaluate the model for your use case before deploying it in production.
License and attribution
DeepSeek-V4.1-Flash is developed by DeepSeek. The NVFP4 checkpoint is produced by NVIDIA. Both are released under the MIT license. Token Factory provides hosted inference for this model. Use of the API is subject to packet.ai's Terms and Privacy Policy.
Model tree for tokenfactorypacketai/DeepSeek-V4.1-Flash-NVFP4
Base model
deepseek-ai/DeepSeek-V4.1-Flash