이 모델의 강점은 무엇인가요?

고급 추론 능력 제공 대규모 컨텍스트 처리 대규모 파라미터 구성

이 모델의 약점은 무엇인가요?

소스 코드 비공개 라이선스 큰 모델 파일 크기 높은 런타임 자원 요구사항

어떤 용도에 가장 적합한가요?

복잡한 논리적 추론 과제 초장문 문서 분석 고급 지식 추출

모델 목록으로

アリババ독점

Qwen3.5-27B

Name: Qwen3.5-27B
Author: アリババ

Qwen3.5-27B는 알리바바가 개발한 추론 모델입니다. 약 2700억 파라미터 규모이며, 101만 토큰의 극도로 긴 컨텍스트 윈도우를 지원합니다.

파라미터

270.0B

컨텍스트

1010K

라이선스

https://huggingface.co/Qwen/Qwen2.5-72B/blob/main/LICENSE

출시일

2026-02-25

API 가격

이 모델의 API 가격 정보는 현재 공개되지 않았습니다

강점

・고급 추론 능력 제공
・대규모 컨텍스트 처리
・대규모 파라미터 구성

약점

・소스 코드 비공개 라이선스
・큰 모델 파일 크기
・높은 런타임 자원 요구사항

활용 사례

・복잡한 논리적 추론 과제
・초장문 문서 분석
・고급 지식 추출

심층 분석

Release Date

February 24, 2026

Parameters

27B

Dense model — all parameters active

Architecture

Hybrid: Gated DeltaNet + Gated Attention (dense)

Context Window

262,144 tokens (native)

Modalities

Text, Image, Video

VRAM (Q4)

~16 GB

VRAM (BF16)

~54 GB

Inference Speed

~35 tok/s on RTX 3090 at Q4

License

Apache 2.0

MMLU-Pro

86.1

강점

・Best creative writing quality in the Qwen3.5 family — denser computation produces more consistent prose
・Strong reasoning: GPQA Diamond 85.5, MMLU-Pro 86.1, IFEval 95.0
・Fits on a single 24GB GPU at Q4 quantization (~16GB VRAM)
・Dense architecture means simpler deployment — no MoE routing complexity
・Natively multimodal with vision and video support

약점

・Slower inference (~35 tok/s) compared to the 35B-A3B MoE model (196 tok/s)
・Lacks the raw speed for batch processing and real-time agent workflows
・Trails the 35B-A3B on throughput-sensitive tasks despite having more active parameters
・Not available as an API model through major providers (primarily self-hosted)
・Quantization at Q4 may impact quality for nuanced creative tasks

경쟁사 비교

Model	Arena	SWE	GPQA	Price
Qwen3.5-35B-A3B	~1390	~65	~83	Open-source
Qwen3.5-9B	~1370	~60	81.7	Open-source
Llama 4 Scout	~1380	~65	~80	Open-source
Qwen3.5-27B	~1400	~68	85.5	Open-source
Mistral Large 2	~1370	~64	~78	Open-source

개요

Qwen3.5-27B is the only dense model in the mid-range of the Qwen3.5 family, offering 27B parameters with all of them active on every token. Released February 24, 2026 under Apache 2.0, it excels at creative writing and complex reasoning where every parameter contributes to output quality. It runs at ~35 tok/s on a single RTX 3090 at Q4, fitting comfortably in 16GB VRAM.

벤치마크 및 성능

Strong across the board: MMLU-Pro 86.1, GPQA Diamond 85.5, IFEval 95.0, C-Eval 92.0. On vision tasks: MMMU 82.3, MathVision 86.0, MathVista 87.8. The dense architecture delivers more consistent quality on nuanced tasks compared to MoE models with similar VRAM requirements. IFEval of 95.0 indicates excellent instruction following — the highest among the Qwen3.5 open models.

상세 비교

Outperforms the 9B on all benchmarks (MMLU-Pro 86.1 vs 82.5, GPQA 85.5 vs 81.7) at the cost of higher VRAM and slower speed. Compared to the 35B-A3B: slower but produces higher quality creative writing and more consistent reasoning. Compared to Llama 4 Scout and Mistral Large 2: competitive on benchmarks with the advantage of native multimodal support and 201 language coverage.

커뮤니티 평가

The 27B occupies a specific niche in the community: users who prioritize quality over speed for creative writing and complex reasoning. It is recommended over the 35B-A3B for novel writing, detailed analysis, and tasks where output consistency matters more than throughput. The community notes it produces 'more consistent prose' than MoE alternatives.

활용 사례

Best for creative writers, researchers, and users who need high-quality nuanced output rather than raw speed. Excellent for long-form content generation, detailed analysis, complex reasoning chains, and tasks where the full 27B active parameters make a qualitative difference. Fits on a single RTX 3090/4090 or 24GB GPU at Q4. Not ideal for batch processing, real-time agents, or speed-critical applications.