🌟 Spotlight

Sovereign AI Foundation Model Project

Nota AI participates in the Sovereign AI Foundation Model Project, a key initiative by the South Korean government (NIPA) to develop global-tier foundation models. As a core optimization partner, we focus on compressing massive LLMs for practical deployment.

🔥 New Release: Solar-Open2-250B Series

Our latest compression of Upstage's flagship Solar-Open2-250B—a 250B-parameter Mixture-of-Experts model—combining routing-aware INT4/NVFP4 quantization with global expert pruning. The result: a frontier-scale MoE that serves on a single B200 (NVFP4) or two H100s (INT4), at near-lossless accuracy. Built on our proprietary DREAM-MoE and SRA-MoE algorithms, which preserve expert-routing decisions under aggressive low-bit quantization.

🏆 Solar-Open2-250B-Nota-INT4-GlobalPruned

The most deployable variant—global expert pruning + INT4 brings the full 250B MoE down to 117.8 GB, running on just 2× H100.

• Global Expert-Sensitivity Pruning: Removes experts using a model-wide importance score (non-uniform, per-layer)—scoring 83.39 average on high-difficulty reasoning benchmarks vs. 78.79 for uniform pruning.
• Near-Lossless: Only −0.6 pt vs. the unpruned INT4 baseline (83.99), ~99% accuracy retention.
• Hardware Efficiency: 117.8 GB weight footprint—serves 131K context on 2× H100 (tensor-parallel), down from the 4× H100 the unpruned INT4 requires.
• Backend Ready: vLLM (W4A16, AutoRound/GPTQ-compatible).

🏆 Solar-Open2-250B-Nota-NVFP4-GlobalPruned

NVFP4 quantization + global expert pruning, tuned for NVIDIA Blackwell—fits the pruned 250B MoE onto a single B200 with 131K context.

• Requires Blackwell (B200/GB200); leverages the FP4 tensor cores unavailable on Hopper, Ada, and Ampere.

Full-Fidelity Quantized Variants (no pruning)

• Solar-Open2-250B-Nota-INT4 — W4A16 quantization, 500.6 GB → 142.9 GB (−71.5%), 81.17 average vs. 81.57 BF16 (−0.5%). vLLM-ready, tensor-parallel across 4+ GPUs.
• Solar-Open2-250B-Nota-NVFP4 — W4A4 for Blackwell, 500.6 GB → 153.3 GB (−69.4%), 99.7% accuracy retention (81.35 vs. 81.57 BF16 average).

🚀 Our Core Business