Alibaba Qwen3.7-Plus: The Multimodal Agent That Sees, Codes, and Thinks

Alibaba’s Qwen team officially released Qwen3.7-Plus on June 2, 2026 — a multimodal agent model that unifies vision and language into a single foundation designed to think, see, and act autonomously . Built on the text backbone of Qwen3.7, the model adds comprehensive vision-language capabilities while preserving full agent intelligence across coding, tool usage, and productivity workflows . The result is a model that can read screenshots, navigate graphical interfaces, write and debug its own code, invoke external APIs, and even iterate on results without human intervention .

Unlike some multimodal models that merely describe images, Qwen3.7-Plus positions itself as a hybrid GUI + CLI agent capable of operating browsers, desktop applications, and terminal environments from visual input alone . One demonstration showed the model completing a real application development lifecycle in 11 hours, generating over 10,000 lines of code end-to-end . It is available through Alibaba Cloud’s Model Studio (Bailian) with API access and a 1-million-token context window .

What Qwen3.7-Plus can actually do

The model’s capabilities go beyond simple image captioning. Qwen3.7-Plus is built around five core agentic abilities :

• Vision-language understanding: Reads and interprets images, video, screenshots, websites, and GUI interfaces — purely for visual understanding, not image generation .
• Deep reasoning: Works through multi-step logical problems that combine visual and textual information .
• Self-programming: Writes, revises, and executes its own code autonomously, allowing it to generate software from visual specifications .
• Tool invocation: Calls external APIs and tools during task execution, making it suitable for real-world automation pipelines .
• Verification and autonomous iteration: Self-checks its outputs and iterates without human feedback — demonstrated by autonomously completing an 11-hour app development cycle .

These capabilities make Qwen3.7-Plus particularly suited for GUI automation, software engineering from screen mockups, browser agents, and cross-framework deployment scenarios (Claude Code, OpenClaw, Qwen Code) . It also maintains the text-agent strengths of the Qwen3.7 family, including competitive coding, reasoning, and instruction-following skills .

Qwen3.7-Plus vs Qwen3.7-Max: Two models, different jobs

Alibaba released the two models simultaneously as previews in May 2026, and they serve distinct purposes despite sharing a common Qwen3.7 foundation .

Capability	Qwen3.7-Max	Qwen3.7-Plus
Modality	Text only	Multimodal — text, images, video, screenshots
Primary use case	Advanced coding, math, long-document reasoning, text-based agents	GUI automation, visual reasoning, hybrid CLI + GUI agents
LM Arena text rank	#13 overall (Elo ~1,475) — #1 among Chinese models	#15 text (some sources report it near Max-level text)
LM Arena vision rank	N/A (cannot process images)	#16 globally — Alibaba ranked #5 lab in vision
Coding rank (Arena)	#10	#12
Key benchmark scores	GPQA Diamond: 92.4%, SWE-bench Verified: 80.4%, Terminal-Bench 2.0: 69.7%, Arena-Hard v2: 90.2	ScreenSpot Pro: 79.0, Terminal-Bench: 70.3 — slightly lower on pure-text metrics but strong on visual agent benchmarks
Pricing (per million tokens)	~$6.00 input / $18.00 output	$0.86–$1.20 input / $2.20–$3.60 output — significantly cheaper

The practical difference: Qwen3.7-Max is the text flagship for developers who need maximum coding and reasoning performance on text-heavy tasks. Qwen3.7-Plus adds vision — the ability to literally see what’s on a screen and act on it — at a much lower price point . If your workflow involves screenshots, browser interfaces, or visual debugging, Max cannot handle those tasks at all; Plus was purpose-built for them .

How the preview performed on LM Arena

Qwen3.7-Max-Preview and Qwen3.7-Plus-Preview entered Arena AI (formerly LMArena) in mid-May 2026, and the rankings made clear each model’s strengths :

• Qwen3.7-Max-Preview: #13 overall in Text Arena (Elo ~1,475), with standout results in specific categories — #7 in Math, #10 in Coding, and #9 in Software/IT. This placed Alibaba as the #6 lab globally in text .
• Qwen3.7-Plus-Preview: #16 overall in Vision Arena, while Alibaba rose to the #5 lab globally in vision. Some Chinese-language sources noted the Plus preview hit as high as the #5 lab ranking in Vision Arena, marking a structural upgrade over Qwen3.6’s vision capabilities .

On independent composite benchmarks, Qwen3.7-Max scored 56.6 on the Artificial Analysis Intelligence Index v4.0 — the highest for a Chinese model at release — and 92/100 on BenchLM.ai, ranking #3 out of 117 models . The Plus model’s standalone benchmark scores remain limited due to its preview status, but early testing shows it holds its own in coding and agent benchmarks while introducing entirely new multimodal capabilities .

Market reaction: Stock surges over 6% on launch day

Investors responded sharply to the Qwen3.7-Plus announcement on June 2, 2026. Alibaba’s shares moved materially on both the Hong Kong and US exchanges :

• Hong Kong-listed shares (9988.HK): Rose as much as 6.84% intraday and closed up 6.6% at HK$130.90 .
• US ADR (BABA): Pre-market trading surged more than 6% to reach $133.23 per share ahead of the US market open .

The immediate catalyst was the formal launch of Qwen3.7-Plus alongside its availability on the Bailian cloud platform, which reinforced Alibaba’s positioning in the rapidly expanding AI agent market . The stock move follows a pattern where previous Qwen model launches — including Qwen3-Max and Qwen3-Coder — have consistently driven positive investor sentiment .

Qwen3.7-Plus represents more than a spec bump. By integrating vision, reasoning, and autonomous execution into a single agent, Alibaba is drawing a line between models that answer questions and models that complete real work across graphical and command-line interfaces. For developers evaluating the Qwen ecosystem, the choice between Max and Plus comes down to one question: do you need a model that reads code, or one that reads your screen?