Model Overview

Description:

The RadixArk Qwen3.8-27B-NVFP4-BF16-LMHead model is a variant of RadixArk/Qwen3.8-27B-NVFP4, the quantized version of Qwen/Qwen3.8-27B produced at RadixArk using NVIDIA Model Optimizer with a mixed NVFP4 W4A4 recipe. The only difference from the source checkpoint is that lm_head is not quantized: it uses the original BF16 lm_head weights from Qwen/Qwen3.8-27B instead of the NVFP4 W4A4 quantized ones. All other tensors are identical to the source checkpoint.

Run on SGLang: launch command and per-platform recipes in the Qwen3.8-27B cookbook.

Third-Party Community Consideration

This model is not owned or developed by RadixArk. It is a quantized derivative of Qwen's model; see the upstream Qwen3.8-27B model card for the source model's capabilities, training information, limitations, and license.

License/Terms of Use:

Apache License 2.0

Deployment Geography:

Global

Use Case:

Developers looking to deploy an off-the-shelf, pre-quantized model in AI agent systems, chatbots, RAG systems, and other AI-powered applications.

Release Date:

Hugging Face 08/14/2026 via https://huggingface.co/RadixArk/Qwen3.8-27B-NVFP4

Model Architecture:

Architecture Type: Transformers (Dense Multimodal)
Network Architecture: Qwen3.8-27B
Number of Model Parameters: 27B

Input:

Input Type(s): Text, image, and video
Input Format(s): String and visual media
Other Properties Related to Input: Native context length up to 262,144 tokens.

Output:

Output Type(s): Text
Output Format: String

Software Integration:

Supported Runtime Engine(s):

• SGLang
  

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Blackwell (this checkpoint was produced and validated on GB300)
  

Preferred Operating System(s):

• Linux
  

Model Version(s):

Quantized with NVIDIA Model Optimizer, commit 87c9f8cf83021957d1a1a575c90c9a4eaaf7ef0c.

Training, Testing, and Evaluation Datasets:

Calibration Data:

Calibration used 1,024 samples from the abisee/cnn_dailymail training split with sequence length 512.

Training Dataset:

RadixArk did not train or fine-tune this checkpoint. Training information is inherited from the upstream Qwen3.8-27B model card.

Evaluation Dataset:

The model was evaluated on GSM8K and Terminal-Bench 2.1.

Post Training Quantization

The MLP gate_proj, up_proj, and down_proj layers use dynamic NVFP4 W4A4 quantization with group size 16. Attention weights use FP8, while lm_head, MTP and vision tensors retain the source BF16 precision.

Usage

The following SGLang configuration uses four NVIDIA Blackwell GPUs:

sglang serve \
  --trust-remote-code \
  --model-path RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead \
  --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

For other deployment topologies and hardware-specific configurations, see the SGLang Qwen3.8-27B cookbook.

Evaluation

The results below were measured on this checkpoint (not inherited from the source NVFP4 checkpoint) on 4x NVIDIA GB300 GPUs with a TP4 SGLang deployment (SGLang main @ d287880a, NEXTN 3/1/4, seed 0).

GSM8K stop rate was 100% with zero truncations and zero errors. For reference, the source NVFP4 checkpoint (quantized lm_head) scored 96.36% (1,271/1,319) on the same protocol; the delta is within single-run sampling noise at temperature 1.0. The reported evaluations were text-only.

Model Limitations:

The base model may generate inaccurate, incomplete, irrelevant, biased, or otherwise undesirable responses. Developers should evaluate the model for their intended use case and apply appropriate safeguards.

Ethical Considerations

RadixArk believes trustworthy AI is a shared responsibility. Developers should ensure that use of this model complies with the upstream license and meets the safety, privacy, and reliability requirements of their application.