모델 목록으로
アリババ오픈소스

Qwen3.5-397B-A17B

Qwen3.5-397B-A17B는 Alibaba에서 개발한 멀티모달 기초 모델입니다. 약 397.0B의 파라미터 규모를 가지며, 최대 256K의 긴 컨텍스트 윈도우를 지원합니다.

파라미터

397.0B

컨텍스트

256K

라이선스

Apache 2.0

출시일

2026-02-16

API 가격

입력 가격 (1M 토큰당)

$0.5

출력 가격 (1M 토큰당)

$

과금 모드: standard

강점

  • 397B의 거대한 파라미터 규모
  • 256K까지의 긴 컨텍스트 이해
  • 매우 다양한 멀티모달 지원

약점

  • 방대한 연산 자원 필요
  • 높은 추론 비용 가능성
  • 모델 크기로 인한 운영 부하

활용 사례

  • 대규모 멀티모달 데이터 분석
  • 초장문 컨텍스트 처리
  • 고급 범용 AI 응용

심층 분석

Release Date

February 16, 2026

Total Parameters

397B

Largest in Qwen3.5 family

Active Parameters

17B per token

MoE with 512 experts, 11 active

Context Window

262,144 tokens (native), up to 1M via YaRN

Architecture

Hybrid MoE: Gated DeltaNet + Gated Attention

Modalities

Text, Image, Video

Natively multimodal

Languages

201

License

Apache 2.0

API Price (Input)

$0.40/1M tokens

API Price (Output)

$2.40/1M tokens

강점

  • Frontier-level open-weight model under Apache 2.0 — best reasoning among open models at launch
  • Natively multimodal: text, image, and video from same weights with no separate VL variant needed
  • Exceptional agentic performance: Terminal-Bench 52.5, BrowseComp 69.0, NOVA-63 59.1
  • 19x faster decoding at 256K tokens vs Qwen3-Max due to hybrid DeltaNet architecture
  • Strong benchmarks: MMLU-Pro 87.8, GPQA Diamond 88.4, SWE-bench 76.4, AIME 2025 91.3

약점

  • Massive hardware requirement: ~220GB VRAM at Q4, ~780GB at BF16 full precision
  • HLE score of 28.7% indicates gaps in absolute expert-level factuality
  • Trails Gemini 3 Pro on competitive coding (LiveCodeBench 83.6 vs 90.7)
  • Occasionally hallucinates tool outputs in autonomous agent scenarios
  • Terminal-Bench 52.5% still leaves room for improvement on complex CLI tasks

경쟁사 비교

ModelArenaSWEGPQAPrice
GPT-5.2~1500~8092.4Proprietary
Claude Opus~149080.987.0Proprietary
Gemini 3 Pro~148076.291.9Proprietary
Qwen3.5-397B-A17B~145076.488.4$0.40/$2.40
DeepSeek V3.2~143073.182.4Proprietary

Qwen3.5-397B-A17B is Alibaba's flagship open-weight model and the largest in the Qwen3.5 family, featuring 397B total parameters with 17B active per token via a hybrid MoE architecture. Released February 16, 2026 under Apache 2.0, it is natively multimodal (text/image/video) with a 262K native context window extendable to 1M. It achieves frontier-competitive benchmarks including GPQA 88.4, SWE-bench 76.4, and AIME 91.3, while running 19x faster at long contexts than its predecessor.

분석 생성일: 2026-05-24