アリババ독점

Qwen3.7-Plus-Preview

Name: Qwen3.7-Plus-Preview
Author: アリババ

이 모델 비교

Alibaba가 개발한 최신 에이전트 특화 모델. 고도의 에이전트 워크플로우와 추론 능력을 구현합니다.

공유:X Hatena LINE

파라미터

Undisclosed

컨텍스트

256K

라이선스

Proprietary

출시일

2026-05-20

일본어 처리 능력

✅High-Quality JP

Multilingual model with strong Japanese language processing capabilities.

API 가격

이 모델의 API 가격 정보는 현재 공개되지 않았습니다

강점

약점

활용 사례

심층 분석

Arena Elo

1460

#15 overall, #12 in coding (BenchLM provisional leaderboard)

SWE-Bench Verified

77.7%

vs Qwen3.7-Max: 80.4%

ScreenSpot Pro (GUI grounding)

79.0

Frontier-tier for computer use agents

Input Price

$0.40/1M tokens

~6x cheaper than Qwen3.7-Max

Context Window

1M tokens

Shared across text, image, and video

Terminal-Bench 2.0

70.3%

Slightly ahead of Qwen3.7-Max (69.7%)

강점

・Exceptional GUI grounding (ScreenSpot Pro 79.0) for computer use and browser automation
・Multimodal interactive hybrid agent unifying vision, code generation, and GUI/CLI operation
・Budget-tier pricing with $0.40/1M input tokens while maintaining near-Max text performance

약점

・Proprietary API-only model with no open weights for self-hosting or customization
・Slightly trails text-only Qwen3.7-Max on pure-text and math benchmarks (e.g., SWE-Bench Verified 77.7% vs 80.4%)
・Visual tokens consume shared 1M context budget, making heavy image/video workloads expensive

경쟁사 비교

Model	Arena	SWE	GPQA	Price
Qwen3.7-Max	1475	80.4%	92.4%	$2.50/$7.50
Claude Sonnet 4.6	1480	~77% (estimated)	~90% (estimated)	$3/$15
DeepSeek V4 Pro (Max)	1485	~80% (estimated)	~91% (estimated)	$1.50/$4.50

개요

Qwen3.7-Plus is Alibaba's flagship multimodal agent model, released in June 2026 as the vision-capable counterpart to the text-only Qwen3.7-Max. It builds on the Qwen3.7 backbone to deliver frontier-tier GUI grounding (ScreenSpot Pro 79.0) and hybrid GUI+CLI agent capabilities, enabling end-to-end automation of tasks like browser navigation, UI recreation, and software development from screenshots. The model uniquely combines vision perception, code generation, and tool use within a single agent loop, allowing it to "see, think, write, act, and verify" in closed-loop execution. Positioned as a "budget multimodal" tier, Plus is roughly six times cheaper than Qwen3.7-Max on input tokens while matching or slightly trailing it on pure-text benchmarks. Its 1M-token context window and 35-hour autonomous run ceiling support long-horizon agent workflows, and it generalizes across agent frameworks like Claude Code, OpenClaw, and Qwen Code. However, it remains API-only with no open weights, which may limit adoption in regulated or air-gapped environments. The model represents Alibaba's strategic move to compete with Western frontier models in the high-stakes GUI agent and computer-use space, offering a cost-effective alternative for teams building screen-aware automation.

벤치마크 및 성능

Qwen3.7-Plus delivers competitive performance across text, multimodal, and agentic benchmarks, approaching Max-tier models while excelling in GUI grounding and hybrid agent tasks. Below is a detailed benchmark comparison from official launch data and independent tracking: | Category | Benchmark | Qwen3.7-Plus | Qwen3.7-Max | Notes | |----------|-----------|--------------|-------------|-------| | **Agentic** | Terminal-Bench 2.0 | 70.3% | 69.7% | Plus slightly ahead | | | OSWorld-Verified | 73.3% | — | GUI automation benchmark | | | AndroidWorld | 81.0% | — | Mobile app operation | | | MCP Atlas | 73.2% | 76.4% | Tool use orchestration | | **Coding** | SWE-bench Verified | 77.7% | 80.4% | Plus trails by 2.7 points | | | SWE-bench Pro | 57.6% | 60.6% | Complex software tasks | | | LiveCodeBench | 89.6% | 91.6% | Code generation | | | SciCode | 51.3% | 53.5% | Scientific coding | | **Reasoning** | MRCRv2 (128K) | 91.7% | 90.4% | Plus ahead | | | CritPt | 9.1% | 13.4% | Max leads on physics | | **Knowledge** | GPQA Diamond | 90.3% | 92.4% | Graduate-level QA | | | HLE | 34.7% | 41.4% | Humanity's Last Exam | | | MMLU-Pro | 88.5% | 89.6% | Multitask understanding | | **Math** | HMMT Feb 2026 | 92.9% | 97.1% | Competition math | | | IMOAnswerBench | 86.0% | 90.0% | Math reasoning | | **Multilingual** | MMLU-ProX (29 langs) | 85.4% | 87% | Multilingual performance | | **Multimodal** | ScreenSpot Pro | 79.0% | — | GUI grounding (frontier) | | | MMMU-Pro | 79% | — | Multimodal understanding | | | CharXiv (RQ) | 85.9% | — | Chart reasoning | | | MathVision | 90.3% | — | Visual math | | | Video-MME (w/ sub.) | 88.0% | — | Video understanding | | **Inst. Following** | IFEval | 94.6% | 94.3% | Near-tie | | | IFBench | 79.1% | 79.1% | Tie | Key takeaways: Plus trails Max on pure-text knowledge (GPQA: 90.3% vs 92.4%) and math (HMMT: 92.9% vs 97.1%), but excels in GUI grounding (ScreenSpot Pro 79.0%) and hybrid agent tasks (Terminal-Bench: 70.3% vs 69.7%). Its multimodal benchmarks are frontier-tier, with strong scores on vision reasoning (MathVision 90.3%) and video understanding (Video-MME 88.0%).

상세 비교

**Qwen3.7-Plus vs Qwen3.7-Max** - **Positioning**: Plus is multimodal with GUI grounding; Max is text-only reasoning flagship. - **Pricing**: Plus is ~6x cheaper ($0.40/$1.60 per 1M tokens vs $2.50/$7.50). - **Performance**: Max leads on text/knowledge benchmarks (GPQA 92.4% vs 90.3%, HLE 41.4% vs 34.7%) and math (HMMT 97.1% vs 92.9%). Plus leads on GUI tasks (ScreenSpot Pro 79.0%) and ties/agentic benchmarks (Terminal-Bench 70.3% vs 69.7%). - **Context**: Both share 1M-token windows and 35-hour autonomous ceilings. - **Use Case**: Choose Max for pure-text coding agents and math-heavy tasks. Choose Plus for browser automation, screenshot-driven workflows, and mixed GUI/CLI agents. **Qwen3.7-Plus vs Claude Sonnet 4.6 (estimated)** - **Positioning**: Both are multimodal agents; Claude emphasizes safety and controllability. - **Pricing**: Plus is cheaper ($0.40/$1.60 vs ~$3/$15 for Claude). - **Performance**: Plus likely leads on GUI grounding (ScreenSpot Pro 79.0% vs ~75% for Claude). Claude may lead on instruction following and safety benchmarks. - **Context**: Plus offers 1M tokens vs Claude's 200K. - **Use Case**: Plus for budget-conscious GUI agents; Claude for safety-critical applications. **Qwen3.7-Plus vs DeepSeek V4 Pro (Max) (estimated)** - **Positioning**: Both are multimodal agents; DeepSeek is known for open-weight models. - **Pricing**: Plus is mid-tier ($0.40/$1.60) vs DeepSeek's aggressive pricing ($1.50/$4.50). - **Performance**: Plus likely leads on GUI grounding; DeepSeek may lead on coding benchmarks. - **Context**: Plus offers 1M tokens vs DeepSeek's 128K. - **Use Case**: Plus for GUI automation; DeepSeek for open-weight flexibility.

커뮤니티 평가

Developer and researcher reactions to Qwen3.7-Plus have been largely positive, particularly praising its GUI grounding capabilities and hybrid agent architecture. On platforms like Twitter and technical blogs, users highlight the ScreenSpot Pro 79.0 score as a breakthrough for non-Western labs in the computer-use space. The model's ability to operate both GUI and CLI environments in a single loop has been noted as a novel architecture that could define future agent products. However, concerns persist around its proprietary nature and ecosystem maturity. Many developers accustomed to Alibaba's open-weight releases (like Qwen3.6) express disappointment at the API-only approach, which limits self-hosting and fine-tuning. Independent verification of benchmark numbers is just beginning, with some early adopters reporting strong performance on browser automation but noting that highly custom UIs may degrade grounding accuracy. Adoption patterns show strong interest from teams building screenshot-to-code tools, browser automation agents, and productivity workflows. The integration with Claude Code, OpenClaw, and Qwen Code has eased adoption for developers already in those ecosystems. Promotional free tiers from Vercel and Command Code have driven initial evaluation, but production deployment is cautious due to the API dependency and context budget limitations for visual-heavy tasks.

활용 사례

**1. Browser Automation & Computer Use** Qwen3.7-Plus excels at navigating web interfaces from screenshots. Example: Given a screenshot of a cloud console, the model can identify form fields, click buttons, and fill in configurations to provision a server. Choose over Max when tasks require visual perception of screens. Choose over Claude Computer Use when budget is primary concern and GUI grounding accuracy (79.0%) is sufficient. **2. UI Recreation & Screenshot-to-Code** The model can generate HTML/CSS/React code from screenshots or mockups. Example: Input a screenshot of a macOS Stocks app; output SwiftUI code with real-time market data integration. Choose over text-only models when visual reference is primary input. Choose over open-weight models when proprietary API access is acceptable. **3. Document & Invoice Processing with Visual Context** Plus can parse images of documents, receipts, or reports, extracting structured data while understanding visual layout. Example: Process a photographed invoice by reading text, tables, and logos, then output JSON. Choose over pure-text models when documents are images rather than digital PDFs. **4. Mobile App Operation from Screenshots** The model can plan and execute mobile UI interactions from Android screenshots. Example: Given a screenshot of a food delivery app, the model can tap buttons, enter addresses, and place orders. Choose over cloud-only agents when mobile emulation is part of the workflow. Note: Requires integration with Appium or similar for execution.