Upstage

Solar Open 2 250B

Name: Solar Open 2 250B
Rating: 88 (10 reviews)
Author: Upstage

最先端

HuggingFace

2.8Kダウンロード546いいねJul 2026公開日1.0M トークンコンテキストUpstage Solarライセンス88 優秀品質

Solar Open 2 250B (250.3000030517578B parameters) requires approximately 154.9 GB of VRAM with Q4_K_M quantization. As a Mixture of Experts model with 15B active parameters, it uses less memory than its total parameter count suggests. For the best balance of quality and speed, we recommend hardware with at least 179 GB of VRAM.

はじめに

— コピー&ペーストでローカル実行

Copy-paste commands to run Solar Open 2 250B on your machine.

Run

docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
  --hf-repo "upstage/Solar-Open2-250B" \
  --hf-file "Solar-Open2-250B-Q4_K_M.gguf" \
  -c 4096 -ngl 99

Quick specs

Parameters250.3B (15B active)

Architecturemoe (MoE)

Context1.0M tokens

Modalitytext

Min RAM97.6 GB

Rec. RAM152.7 GB (Q4_K_M)

LicenseUpstage Solar

FamilySolar

✓ Code✓ Reasoning

About this model

Solar Open 2 is Upstage's 250B-A15B open-weight Hybrid-Attention Mixture-of-Experts model, built for agentic work such as office productivity, document-heavy tasks and coding. It interleaves three linear-attention layers with one softmax-attention layer, so only 12 of its 48 layers keep a KV cache — holding long-context memory to roughly a quarter of an all-softmax model of the same shape.

•250B total / 15B active per token — 320 routed experts plus 1 shared expert, top-8.
•1M-token context; linear-attention layers encode order intrinsically, so RoPE is removed entirely (NoPE).
•Only 12 of 48 layers keep a KV cache — roughly a quarter the long-context memory of an all-softmax model.
•Purpose-built for agentic workflows: tool calling, multi-step reasoning, end-to-end task execution.
•Trilingual: English, Korean and Japanese.

Solar Open 2 250Bのおすすめ

AMD Instinct MI325X 256GBS

256 GB

AMD Instinct MI350X 288GBS

このモデルを実行

Solar Open 2 250B on AMD Instinct MI325X 256GB Solar Open 2 250B on AMD Instinct MI350X 288GB Solar Open 2 250B on NVIDIA B200 180GB

量子化オプション

量子化レベル別VRAM推定値

No hardware detected — fit column shows raw VRAM estimates

Quant	Bits	VRAM	Quality	Fit
Q1_0_G128	1.125	36.0 GB	Very Low	—
Q2_0_G128	1.71	66.8 GB	Low	—
Q2_K	2	97.6 GB	Low	—
Q3_K_S	3	122.6 GB	Low	—
NVFP4	4	140.2 GB	Medium	—
Q4_K_M	4	152.7 GB	Medium	—
Q5_K_M	5	180.2 GB	High	—
Q6_K	6	205.2 GB	High	—
Q8_0	8	267.8 GB	Very High	—
F16	16	513.1 GB	Maximum	—

Quality benchmarks

Solar Open 2 250B benchmark scores

Benchmark verified

Coding

SWE-bench Verified70.4%

HumanEval+—

Aider Polyglot—

LiveCodeBench92.4%

Reasoning

MMLU-Pro86.2%

GPQA Diamond86.3%

MATH-500—

ARC Challenge—

Source: official · 2026-07-22

ハードウェア互換性

全ハードウェアの適合度推定

カリキュレーターを開く

Computing compatibility...

メモリ内訳

Reference: RTX 2060 6GB

Weights152.7 GB

KV Cache0.7 GB

Runtime0.9 GB

Headroom0.6 GB

よくある質問