nota-ai/Solar-Open2-250B-Nota-INT4

Solar Open2 250B — Nota INT4

Nota AI presents a 4-bit quantized release of Upstage's Solar Open2 250B, produced with Nota AI's proprietary quantization technology specialized for Mixture-of-Experts (MoE) large language models.

Highlights

• W4A16 (weight-only INT4) — group_size=128, packed in the auto_round (GPTQ-compatible) format. Formatted with AutoRound for direct, out-of-the-box serving in vLLM.
• Nota AI's proprietary MoE quantization framework. This release is built upon a suite of techniques developed by Nota AI to preserve model quality under aggressive low-bit quantization of MoE architectures:
  • A MoE-specialized calibration-dataset construction method, which achieved 1st place across all tracks at the NVIDIA Nemotron Hackathon.
  • DREAM-MoE and SRA-MoE, two quantization algorithms proposed by Nota AI (published at the ICML 2026 Workshop on AdaptFM), which preserve MoE routing decisions and align expert-routing behavior throughout the quantization process.

License

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

1. MODEL WEIGHTS (*.safetensors) The Upstage Solar License is based on the permissive Apache License 2.0 with specific mandatory conditions to support the Solar ecosystem.

   Key requirements for Derivative AI Models (create / train / fine-tune / distill / improve using Solar Open 2):

   • Naming: prefix your model name with "Solar" (e.g., Solar-MyModel-v1).
   • Attribution: prominently display "Built with Solar" in related public-facing materials.
   • Notice: include a copy of the Upstage Solar License with your derivative model.

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

Performance

Weight footprint

Precision	Weight footprint
BF16	500.6 GB
Nota INT4.	142.9 GB

Benchmarks

Benchmark	BF16	Nota INT4
Tau2-Bench	75.20	74.04
HLE	27.88	27.15
GPQA Diamond	86.26	85.86
IFBench	80.00	80.54
LiveCodeBench (v5–v6)	87.03	85.50
MMLU-Pro	86.19	86.29
AIME 2026 (EN)	95.67	96.00
IFEval (EN)	94.09	93.35
HMMT	92.05	90.91
KMMLU-Pro	78.38	78.79
HAE-RAE Bench v1.1	73.84	72.53
AIME (KO)	97.67	97.67
KBL	75.51	75.20
KBank-MMLU	80.80	80.88
KorMedMCQA	92.99	92.85
Avg.	81.57	81.17

Usage

This model is packed in the AutoRound (GPTQ-compatible) INT4 format and can be served directly with vLLM:

vllm serve nota-ai/Solar-Open2-250B-Nota-INT4 \
  --tensor-parallel-size 4 \
  --trust-remote-code

• Set --tensor-parallel-size according to the number of GPUs available in your serving environment.

• For reasoning and tool-calling, use the reasoning parser and tool parser provided in the original upstage/Solar-Open2-250B model card.

• See the original model card for the prompt format, parser configuration, and further details.

Citation

@inproceedings{park2026dreammoe,
  title     = {{DREAM-MoE}: Downstream Routing Error-Aware Margin-Preserving Quantization for Mixture-of-Experts Large Language Models},
  author    = {Park, Hancheol and Lee, Geonho and Kim, Tae-Ho},
  booktitle = {ICML 2026 Workshop on Resource-Adaptive Foundation Model Inference (AdaptFM)},
  year      = {2026},
  url       = {https://openreview.net/forum?id=Wyhqwjl51A},
}

@inproceedings{lee2026sramoe,
  title     = {{SRA-MoE}: Output-Aware Selective Router Alignment for MoE Quantization},
  author    = {Lee, Geonho and Park, Hancheol and Lee, Seunghyun and Choi, Jungwook and Kim, Tae-Ho},
  booktitle = {ICML 2026 Workshop on Resource-Adaptive Foundation Model Inference (AdaptFM)},
  year      = {2026},
  url       = {https://openreview.net/forum?id=H0NoX02erJ},
}