이 모델의 강점은 무엇인가요?

1조 2천억 이상의 파라미터로 방대한 용량 262K 극장문 컨텍스트 이해 고급 추론 능력 달성

이 모델의 약점은 무엇인가요?

소스 코드 비공개 라이선스 모델 거대한 모델 규모로 인한 높은 연산 비용 사용 제한이 있는 라이선스

어떤 용도에 가장 적합한가요?

복잡한 논리적 추론 과제 초장문 문서 분석 고급 지식 집약적 과제

모델 목록으로

アリババ독점

Qwen3.5-122B-A10B

Name: Qwen3.5-122B-A10B
Author: アリババ

Qwen3.5-122B-A10B는 알리바바가 개발한 추론 모델입니다. 약 1조 2200억 파라미터의 대규모 매개변수와 26만 2천 토큰의 긴 컨텍스트 윈도우를 특징으로 합니다.

파라미터

1220.0B

컨텍스트

262K

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

・1조 2천억 이상의 파라미터로 방대한 용량
・262K 극장문 컨텍스트 이해
・고급 추론 능력 달성

약점

・소스 코드 비공개 라이선스 모델
・거대한 모델 규모로 인한 높은 연산 비용
・사용 제한이 있는 라이선스

활용 사례

・복잡한 논리적 추론 과제
・초장문 문서 분석
・고급 지식 집약적 과제

심층 분석

Release Date

February 2026

Total Parameters

122B

MoE architecture

Active Parameters

10B per token

256 experts with selective activation

Context Window

262,144 tokens

Architecture

Hybrid MoE: Gated DeltaNet + Gated Attention

Modalities

Text, Image, Video

VRAM (Q4)

~70 GB

VRAM (BF16)

~244 GB

License

Apache 2.0

API Price

Available via DashScope, SiliconFlow, DeepInfra

강점

・Strong quality-to-compute ratio: 10B active parameters deliver near-frontier performance
・Natively multimodal with text, image, and video support from the same weights
・Fits on multi-GPU consumer setups at Q4 (~70GB VRAM) — accessible for serious enthusiasts
・262K native context with hybrid DeltaNet architecture for fast long-context inference
・Apache 2.0 license enables commercial use and fine-tuning

약점

・Positioned awkwardly between the 397B flagship and the 35B-A3B speed model
・70GB VRAM at Q4 still requires multi-GPU setup (2x RTX 4090 or better)
・No dedicated benchmark spotlight — overshadowed by the 397B and 9B in marketing
・Active parameters (10B) may be insufficient for the most demanding reasoning tasks
・Community adoption has been slower compared to the 9B, 27B, and 35B-A3B

경쟁사 비교

Model	Arena	SWE	GPQA	Price
Qwen3.5-397B-A17B	~1450	76.4	88.4	$0.40/$2.40
Qwen3.5-27B	~1400	~68	85.5	Open-source
Llama 4 Scout	~1380	~65	~80	Open-source
Qwen3.5-122B-A10B	~1420	~72	~86	Open-source
DeepSeek V3	~1410	~70	~82	$0.14/$0.28

개요

Qwen3.5-122B-A10B is the mid-tier MoE model in the Qwen3.5 family, with 122B total parameters and 10B active per token. It provides a strong balance between quality and inference efficiency, fitting on multi-GPU setups at ~70GB VRAM in Q4 quantization. Released under Apache 2.0 in February 2026, it is natively multimodal and benefits from the same hybrid DeltaNet architecture as its larger sibling.

벤치마크 및 성능

Benchmark data from the Qwen3.5 family shows the 122B-A10B positioned between the 397B and 27B dense model. On MMLU-Pro it scores in the ~85 range, GPQA Diamond ~86, with strong multimodal performance from the shared vision encoder. It excels on instruction following and document understanding tasks. The MoE design with 256 experts and 10 active provides strong specialization capacity while keeping compute costs reasonable.

상세 비교

Occupies the middle ground in the Qwen3.5 lineup. Delivers roughly 90-95% of the 397B's quality at a fraction of the hardware cost. Outperforms the 27B dense model on most benchmarks while requiring similar or less VRAM at Q4 quantization (70GB vs 16GB for 27B, but significantly more capable). Compared to Llama 4 Scout and DeepSeek V3, it offers competitive quality with native multimodal support.

커뮤니티 평가

The 122B-A10B has received less community attention than the headline 397B or the value-star 9B/35B-A3B models. It is valued by users with multi-GPU setups who want strong quality without the extreme VRAM demands of the 397B. SiliconFlow and DeepInfra offer hosted API access, making it accessible without local hardware.

활용 사례

Best for teams and researchers with multi-GPU servers (2-4 GPUs) who need strong general-purpose AI with multimodal capabilities. Good balance for production workloads where the 397B is too expensive to host but the 27B is insufficient. Suitable for document analysis, image understanding, coding assistance, and long-context processing. API access through DashScope and SiliconFlow makes it viable for users without local hardware.