Solar Open 2 model card · inference-research

Fields

id

sources/solar-open2-250b-model-card

URL

https://huggingface.co/upstage/Solar-Open2-250B

Publisher

Upstage

Kind

repo

Published

2026-07-22

Accessed

2026-10-18

notes

Card read 2026-10-18. Hybrid-attention MoE, nameplate 250B-A15B, 15B activated per token, 320 routed plus 1 shared expert, 8 routed per token, 48 layers of which 12 are softmax and 36 linear, NoPE, 1M context. HF API createdAt 2026-07-22T02:40:42Z, license other (Upstage Solar License), safetensors BF16 250,287,794,944 plus F32 15,360. The English benchmark table reports LiveCodeBench v6 92.4, MMLU-Pro 86.2, GPQA Diamond 86.3, AIME 2026 95.7, SWE-bench Verified 70.4, IFBench 80.0. Vendor claims, not re-run. Technical report arXiv 2607.20062.

Notes

Card read 2026-10-18. Hybrid-attention MoE, nameplate 250B-A15B, 15B activated per token, 320 routed plus 1 shared expert, 8 routed per token, 48 layers of which 12 are softmax and 36 linear, NoPE, 1M context. HF API createdAt 2026-07-22T02:40:42Z, license other (Upstage Solar License), safetensors BF16 250,287,794,944 plus F32 15,360. The English benchmark table reports LiveCodeBench v6 92.4, MMLU-Pro 86.2, GPQA Diamond 86.3, AIME 2026 95.7, SWE-bench Verified 70.4, IFBench 80.0. Vendor claims, not re-run. Technical report arXiv 2607.20062.

Cross-references

This record references no other records.