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
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| GPT-5.2 | ~1500 | ~80 | 92.4 | Proprietary |
| Claude Opus | ~1490 | 80.9 | 87.0 | Proprietary |
| Gemini 3 Pro | ~1480 | 76.2 | 91.9 | Proprietary |
| Qwen3.5-397B-A17B | ~1450 | 76.4 | 88.4 | $0.40/$2.40 |
| DeepSeek V3.2 | ~1430 | 73.1 | 82.4 | Proprietary |
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.
출처
분석 생성일: 2026-05-24