Solar-Open-100B-NotaMoeQuant-NVFP4

This repository provides Upstage’s flagship model, Solar-Open-100B, packaged with Nota AI’s proprietary quantization technique specifically developed for Mixture-of-Experts (MoE)-based LLMs. Unlike conventional quantization methods, this approach incorporates a novel method designed to mitigate representation distortion that can occur when experts are mixed under quantization in MoE architectures.

Overview

• Base model: Solar-Open-100B
• Quantization: NVFP4
• Packing format: compressed-tensors (ensuring backend compatibility with HF and vLLM)
• Hardware Requirements:
  • Minimum: 1 x NVIDIA B100
  • We have tested on B100, B200, and B300.

License

This repository contains both model weights and code, which are licensed under different terms:

1. MODEL WEIGHTS (*.safetensors) Licensed under Upstage Solar License See: https://huggingface.co/upstage/Solar-Open-100B/blob/main/LICENSE

2. CODE (*.py, *.json, *.jinja files) Licensed under Apache License 2.0 See: https://www.apache.org/licenses/LICENSE-2.0

Performance

• English

	Solar-Open-100B	Nota MoE Quantization (Ours)	AutoRound
PPL (WikiText-2)↓	6.06	6.90	7.22
MMLU-Pro↑	73.91	62.53	61.56
GPQA-Diamond↑	58.08	45.96	42.42
General Evaluation Benchmarks	75.77	73.94	73.74

• Model weigth memory footprint

Solar-Open-100B	Nota MoE Quantization (Ours)
191.2 GB	58.7 GB

• Note
  • General evaluation benchmarks: relatively low-difficulty tasks that typically require short responses (ARC-C, ARC-E, BoolQ, HellaSwag, MMLU, PIQA, TruthfulQA, WinoGrande, GSM8K). The score is calculated by averaging across all tasks.
  • ↑ / ↓ denote the direction of improvement: higher is better (↑), lower is better (↓).
  • Because we used a smaller thinking budget (8,192 tokens), the results for MMLU-Pro and GPQA-Diamond are slightly lower than the numbers reported in the original Solar-Open-100B repository.
  • Memory refers to the pure VRAM footprint occupied only by the model weights.

Inference

vLLM

Step 1: Create and activate a Python virtual environment

uv venv --python 3.12 --seed
source .venv/bin/activate

Step 2: Install Solar Open's optimized vLLM

pip install vllm==0.17.0

Step 3: Overwrite the two files (solar_open.py and registry.py) in the patches folder of the repository containing the model weights into the vllm/model_executor/models directory inside the folder where vLLM is installed (typically lib/python3.xx/site-packages).

Step 4: Start the vLLM server (For 1GPUs)

vllm serve nota-ai/Solar-Open-100B-NotaMoEQuant-NVFP4 \
    --served-model-name Solar-Open \
    --trust-remote-code \
    --tensor-parallel-size 1 

Step 5: Generate the response

from openai import OpenAI

client = OpenAI(
    base_url="http://0.0.0.0:8000/v1",
    api_key="EMPTY"
)

response = client.chat.completions.create(
    model="Solar-Open",
    messages=[
        {"role": "user", "content": "who are you?"}
    ],
    temperature=0.8,
    top_p=0.95,
)

print(response.choices[0].message.content)