Qwen3.8-27B-NVFP4 by RadixArk

Qwen3.8-27B-NVFP4 by RadixArk

A pre-quantized NVFP4 copy of Qwen3.8-27B that reads text, images and video and answers in text, served with SGLang on NVIDIA Blackwell GPUs. Suits developers running agents, chatbots and RAG on that hardware.

Made by RadixArk on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a quantized copy of Qwen3.8-27B made by RadixArk, using NVIDIA Model Optimizer and a mixed NVFP4 W4A4 recipe. It is a multimodal model that takes text, images and video and answers in text. RadixArk did not train it, so its abilities come from the original Qwen model.

Who it suits: Developers who want a ready-quantized model for agents, chatbots and retrieval systems, and who have NVIDIA Blackwell GPUs. The card tests it on four GPUs, so this is a setup for servers, not a laptop.

Industry | Who uses it | How they use it | Example
AI applications | Agent developers | Serve a pre-quantized model behind an agent | Run it with SGLang and call its tool parser
Software | Chatbot teams | Host a large multimodal chat model | Serve text, image and video questions
Search and RAG | Retrieval engineers | Use it as the answering model in a RAG system | Answer from retrieved documents
Infrastructure | Inference engineers | Test NVFP4 on Blackwell hardware | Measure speed on four GPUs

Anyone without NVIDIA Blackwell GPUs can skip it, since the card names that hardware and the SGLang engine only.

Where to use it:

• SGLang, the one runtime engine the card lists
• NVIDIA Blackwell GPUs, with the checkpoint produced and validated on GB300
• Linux, the preferred operating system

Use case | Who | What they get back
Agent backend | Agent developers | Text answers with reasoning and tool calls parsed by SGLang
Chatbot serving | Product teams | Replies to text, image and video input
RAG answering | Search teams | Answers over long inputs, with context up to 262,144 tokens

How to set it up:

• Python with pip, to download the files
• SGLang installed on a Linux machine
• NVIDIA Blackwell GPUs, four in the card's example, and about 18.2 GB of weights to store

1. Install the Hugging Face hub package.
2. Download the model files.
3. Launch SGLang with the command below, which the card gives for four Blackwell GPUs.

pip install -U huggingface_hub
hf download RadixArk/Qwen3.8-27B-NVFP4

How to use it: Start the SGLang server with this configuration, then send chat requests to it. The card gives it for four NVIDIA Blackwell GPUs and points to a cookbook for other layouts.

sglang serve \
  --trust-remote-code \
  --model-path RadixArk/Qwen3.8-27B-NVFP4 \
  --tp-size 4 \
  --mem-fraction-static 0.75 \
  --speculative-algorithm NEXTN \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --host 0.0.0.0 \
  --port 30000

Tool | What it does
sglang serve | Starts the server for the model
--tp-size 4 | Splits the model across four GPUs
--reasoning-parser qwen3 | Separates the model's reasoning from its answer
--tool-call-parser qwen3_coder | Reads tool calls out of the output

Example: You start the server with the command above and send it a chat request that includes a screenshot and a question. The server returns the model's text answer, with any reasoning separated by the parser.

For newcomers and experts: If you are new, know that this is a server-class model and the first step is simply to get the SGLang command running on suitable GPUs. Check the plain text replies before you add tools or images.

Experts get a dense multimodal Transformer based on Qwen3.8-27B. Dynamic NVFP4 W4A4 quantization covers the MLP projection layers and the output head, in groups of 16. Attention weights are stored in FP8, and the MTP and vision parts keep their original BF16 precision. Calibration used 1,024 samples from the cnn_dailymail training split at length 512.

Title: Reported scores on this NVFP4 checkpoint
GSM8K: 97.27 percent
Terminal-Bench 2.1: 73.81 percent

Download the NVFP4 checkpoint
Launch SGLang on Blackwell GPUs
Send text, image or video requests
Read the text answer

Good to know:

• Scores: GSM8K was 1,283 of 1,319 in thinking mode, and Terminal-Bench 2.1 was 62 of 84 tasks, on four B300 or GB300 GPUs
• Evaluation scope: the reported tests were text-only
• Limits: the base model may give inaccurate, biased or unwanted answers, so test it for your use
• License: Apache 2.0
• Cost to run: none named, no API key or account, but it needs NVIDIA Blackwell hardware
• Data it reaches: it downloads files from Hugging Face and reads the text, images and video you send
• Not for: the card says it is a third-party quantized derivative, so check the upstream Qwen card for the source model's limits

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: Model Optimizer
• Parameters: 18.2B
• Weights: about 18.2 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 979,445
• Likes: 126
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 2
• Fine-tunes: 4
• Tags: safetensors, qwen3_5, radixark, modelopt, qwen3.8, quantized, fp4, nvfp4, image-text-to-text, conversational, 8-bit

Links

• Model: https://huggingface.co/RadixArk/Qwen3.8-27B-NVFP4
• Files: https://huggingface.co/RadixArk/Qwen3.8-27B-NVFP4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download RadixArk/Qwen3.8-27B-NVFP4

Then follow the usage in its model card. Nothing in this kit runs by itself.