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

NVFP4 (4-bit float, W4A4) — group_size=16, packed in the llm-compressor (compressed-tensors) format for direct serving in vLLM. Both weights and activations are quantized to 4-bit floating point.
- Requires NVIDIA Blackwell. NVFP4 relies on the FP4 tensor cores introduced in the Blackwell architecture (e.g. B200 / GB200), so inference must run on a Blackwell-class GPU. Earlier architectures (Hopper, Ada, Ampere) do not support NVFP4 execution.
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

Solar Open 2 is distributed under the Upstage Solar License.

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.

Performance

Weight footprint

Precision	Weight footprint
BF16	500.6 GB
Nota NVFP4	153.3 GB

Benchmarks

Benchmark	BF16	Nota NVFP4
Tau2-Bench	75.20	75.08
HLE	27.88	27.66
GPQA Diamond	86.26	85.45
IFBench	80.00	81.02
LiveCodeBench (v5–v6)	87.03	88.55
MMLU-Pro	86.19	86.15
AIME 2026 (EN)	95.67	96.67
IFEval (EN)	94.09	92.61
HMMT	92.05	90.15
KMMLU-Pro	78.38	77.93
HAE-RAE Bench v1.1	73.84	72.84
AIME (KO)	97.67	97.00
KBL	75.51	75.40
KBank-MMLU	80.80	80.68
KorMedMCQA	92.99	93.05
Avg.	81.57	81.35

Latency

Measured with vLLM v0.1.dev16959+g00907fc9b on 4×NVIDIA B300 SXM6 (TP=4). Long-context serving workload: 32,768 input tokens / 2,048 output tokens per request. All latency figures are p50 (median).

Concurrency = 1 (single user — latency floor)

Model	Output tok/s	TTFT (ms)	TPOT (ms)	tok/s/user	E2E latency (s)
Solar-Open-2 (BF16)	119.8	975	7.87	127.0	17.1
Solar-Open-2-Nota-NVFP4	144.5	734	6.56	152.5	14.2
Solar-Open-2-Nota-NVFP4-Global-Pruned	146.9	741	6.45	155.1	13.9

Speedup vs BF16 — NVFP4 1.21×, NVFP4-Global-Pruned 1.23× (output throughput).

Concurrency = 32 (32 simultaneous users)

Model	Output tok/s (all 32 users)	Total tok/s (all 32 users)	Req/s	tok/s/user (per user)	TTFT (ms)	TPOT (ms)	E2E latency (s)
Solar-Open-2 (BF16)	1,010.8	17,183.6	0.49	35.2	4,856	28.44	62.4
Solar-Open-2-Nota-NVFP4	1,468.2	24,958.6	0.72	53.7	4,893	18.62	42.4
Solar-Open-2-Nota-NVFP4-Global-Pruned	1,503.5	25,560.2	0.73	56.2	5,337	17.79	41.6

Speedup vs BF16 — NVFP4 1.45×, NVFP4-Global-Pruned 1.49× (output throughput).

Reading the numbers. Output tok/s and Total tok/s are aggregate across all concurrent users (server capacity). tok/s/user, TTFT, TPOT, and E2E latency are what a single user experiences (tok/s/user = 1000 / TPOT).

Quick Start

This model is packed in the NVFP4 (compressed-tensors) format and can be served directly with vLLM on a Blackwell-class GPU:

bash

uv venv --python 3.12 --seed solar_open2_venv
source .venv/bin/activate

VLLM_PRECOMPILED_WHEEL_LOCATION="https://github.com/vllm-project/vllm/releases/download/v0.22.0/vllm-0.22.0%2Bcu129-cp38-abi3-manylinux_2_28_x86_64.whl" \
VLLM_USE_PRECOMPILED=1 \
uv pip install --reinstall-package vllm --torch-backend=cu129 \
  "git+https://github.com/UpstageAI/vllm.git@v0.22.0-solar-open2"

bash

vllm serve nota-ai/Solar-Open2-250B-Nota-NVFP4 \
  --served-model-name solar-open2-250b \
  --tensor-parallel-size 4 \
  --default-chat-template-kwargs '{"think_render_option":"preserved"}' \
  --reasoning-parser solar_open2 \
  --tool-call-parser solar_open2 \
  --enable-auto-tool-choice \
  --logits-processors vllm.v1.sample.logits_processor.solar_open2:SolarOpen2TemplateLogitsProcessor

Set --tensor-parallel-size according to the number of GPUs available in your serving environment.
See the original model card for the prompt format, parser configuration, and further details.

Send a chat completion request:

bash

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "solar-open2-250b",
    "messages": [
      {"role": "user", "content": "What is Upstage?"}
    ],
    "max_tokens": 131584,
    "temperature": 1.0,
    "top_p": 1.0,
    "reasoning_effort": "high"
  }'

Citation

bibtex

@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},
}