Alibaba’s Qwen3.7-Plus Tops GUI Vision and Goes Closed-Source

Alibaba’s Qwen3.7-Plus reached general availability on June 1, scoring 79.0 on ScreenSpot Pro (a graphical user interface (GUI) grounding benchmark) and pricing at $0.40 per million input tokens. The model accepts text, images, and video as input, runs a one-million-token context window, and pairs with the text-only Qwen3.7-Max introduced at the Alibaba Cloud Summit in Hangzhou on May 20.

Qwen3.7-Plus is the fourth major Qwen release to ship without open weights since March, following Qwen3.5-Omni, Qwen3.6-Plus, and Qwen3.7-Max. The lab had previously released over 300 models under open licenses.

Vision Scores That Put Alibaba in the GUI Race

ScreenSpot Pro tests one specific capability: given a screenshot and a natural-language instruction, can the model identify the exact UI element or pixel to click? It’s the operative benchmark for any agent that needs to drive browsers, navigate desktop software, or fill forms from visual input alone. The Qwen team’s announcement on X described Qwen3.7-Plus as a “multimodal interactive hybrid agent” that perceives real-world scenes, reads and operates graphical interfaces, and writes code from visual references.

• 79.0: Qwen3.7-Plus on ScreenSpot Pro, ahead of GPT-5.4 at 67.4 and Claude Opus 4.6 at 49.5
• 70.3: Qwen3.7-Plus on Terminal-Bench 2.0-Terminus (safe, iterative terminal code execution), beating DeepSeek-V4-Pro Max at 67.9 and Gemini-3.1 Pro at 63.5
• #16 overall on Vision Arena (LM Arena’s image-understanding leaderboard), putting Alibaba as the fifth-ranked lab in vision globally

Terminal-Bench 2.0-Terminus tests whether a model can safely execute real shell commands across multi-step loops, a capability distinct from visual grounding. Qwen3.7-Max, text-only with no vision parameters at all, scores 69.7 on that benchmark. Qwen3.7-Plus scores 70.3, narrowly ahead of the text-only variant despite carrying additional vision parameters that compete for network capacity, which matters for teams trying to run a single model across both GUI navigation and CLI tasks rather than maintaining two separate endpoints.

The Qwen team ran a public demonstration: a Qwen3.7-Plus-powered agent built an English vocabulary learning application autonomously, running for over eleven hours, producing more than 10,000 lines of code across more than 1,000 agent calls. Requirements documentation, automated code generation, GUI-based testing, and independent version management all ran within the same loop. A preview of the model appeared on LM Arena around May 14, giving developers roughly 18 days of production signal before the GA endpoint dropped the preview suffix on June 1.

The Pricing Math Against the Frontier

At $0.40 per million input tokens and $1.60 per million output, Qwen3.7-Plus prices well below the frontier multimodal API (application programming interface) options from Western labs. For agent pipelines that repeatedly read the same codebase or UI kit across hundreds of loops, Alibaba offers cached inputs at $0.04 per million tokens, a tenth of the standard rate.

Qwen 3.7 Plus being 40% cheaper than Max changes the conversation. If the output is close enough for most coding and much stronger for visual workflows, do you really need Max every day or only for the heavy terminal-only jobs?

@Boxmining, a Web3 venture capitalist and prominent industry voice, posted the question to X shortly after the model’s launch.

The Chinese competition is close on price. MiniMax-M3, which also offers multimodal capability, prices input at $0.30 per million and output at $1.20. DeepSeek-V4-Pro undercuts on output at $0.87 per million but is a text-and-code model without the GUI-grounding architecture central to what Qwen3.7-Plus is designed to do.

A Lab That Open-Sourced Over 300 Models and Then Stopped

The Departures That Preceded the Turn

The Qwen family, first open-sourced in August 2023, grew into the most downloaded AI model ecosystem on Hugging Face. Alibaba’s 2025 “Change the World” recognition from Fortune cited over 300 generative AI models released under open licenses, and Fortune editors wrote that the strategy had pushed U.S. competitors to release their own open-source models. Stanford University and the University of Washington built top-performing derivative models for under $50 using Qwen weights; UC Berkeley trained a reinforcement-learning math model for under $30.

Then the lab’s leadership changed. Lin Junyang, Qwen’s technical lead, resigned in March 2026, the third senior departure from the AI unit that year. Hao Zhou, a veteran of Google DeepMind’s Gemini program, took over, bringing a development culture that has not publicly released Gemini’s frontier-tier model weights.

MaaS, Token Hub, and the New Mandate

CEO Eddie Wu named the new direction “MaaS” (model as a service) around the same time: route inference through Alibaba Cloud, bill per token, retain frontier capabilities as a commercial service. The company created Alibaba Token Hub (ATH), consolidating five AI units under Wu’s direct supervision.

Qwen3.5-Omni launched March 30, 2026, as a proprietary API-only model. Qwen3.6-Plus followed in April under the same restriction. Qwen3.7-Max launched May 20 without open weights. Qwen3.7-Plus arrived June 1 the same way. One model breaks from the sequence: Qwen3.6-35B-A3B, released in April 2026, still carries Apache 2.0 weights on Hugging Face and remains free to download. The pattern across both generations is consistent: smaller efficiency variants stay open, multimodal flagships and top-tier frontiers do not.

The 290,000 Developers Who Built on Free Weights

Open Qwen weights under Apache 2.0 meant no per-query billing. An enterprise running a prior Qwen model on its own GPU cluster amortized the hardware cost once and ran unlimited queries at compute cost only. Steve Frey, an AI industry cofounder and product lead, has estimated that self-hosting via open weights cuts costs 5-10x compared to proprietary API rates, driven by the efficiency of the mixture-of-experts architectures the Qwen family uses.

A marketing agency processing 5,000 ad images per week on a self-hosted Qwen3-VL instance was paying roughly $0.002 per image, against $0.02 to $0.05 for comparable proprietary vision APIs. That same agency would face a 10x to 25x cost increase migrating to any API-only alternative. Qwen3.7-Plus at API rates does not replicate those self-hosted economics, and the self-hosted path is not available for the current generation.

The community that relied on free weights includes:

• Over 290,000 developers worldwide who built on open Qwen weights under Apache 2.0
• More than 113,000 fine-tuned model variants on Hugging Face, derived from free community access to prior Qwen releases
• Airbnb’s customer-service chatbot, confirmed by CEO Brian Chesky to run on Qwen
• Korean startup Univa, which cut its AI costs 30% by self-hosting Qwen instead of paying for proprietary alternatives

Alibaba has not stated a public timeline for releasing open weights for any of its 3.7-generation models.

Preserve Thinking and the Stateful Agent Problem

The preserve_thinking parameter, exposed in the API, retains internal reasoning blocks (marked as <think> tokens) across consecutive turns. In a multi-step agent workflow, this addresses a specific failure mode: a coding agent that spends steps one through ten decomposing a dependency graph can reach step 30 and still hold the original rationale, without a full recomputation mid-task.

Alibaba introduced the feature during the Qwen 3.6 generation. The Qwen3.6 GitHub repository, which hosts the open Apache 2.0 weights, carries it alongside the proprietary variants, so the parameter is a design choice baked into the generation rather than a feature reserved for paid endpoints. Anthropic calls an equivalent capability “Extended Thinking” for Claude Opus 4.8; OpenAI implements a similar mechanism as encrypted reasoning pass-back in GPT-5.5, requiring developers to return reasoning items from prior function calls.

The model allocates up to 256,000 tokens specifically for internal chain-of-thought processing within its one-million-token context window. For an automated cloud migration agent reading an entire codebase before touching a single file, that allocation holds the full dependency map and edge-case analysis at the same time.

Dunjie Lu, a research intern at Alibaba Qwen, noted on X that Qwen3.7-Plus “shows clear gains over Qwen3.6-Plus in computer-use capabilities, with stronger generalization beyond general desktop tasks into professional workflows such as data engineering and scientific research.”

What the Sovereignty Clause Rules Out

Qwen3.7-Plus is served exclusively through Alibaba Cloud Model Studio. All inference passes through the company’s international endpoints, with developer documentation pointing to the Singapore instance. The weights cannot be downloaded, and there is no on-premises or air-gapped deployment path.

For healthcare companies under HIPAA (the U.S. federal health privacy law), European enterprises with GDPR (the EU’s data-residency regulation) obligations, and defense contractors under export-control requirements, the external cloud routing is a compliance question that precedes any capability or pricing evaluation. The 79.0 on ScreenSpot Pro and the $0.40 per million input rate are irrelevant if the data cannot legally leave internal infrastructure.

Alibaba has offered no public roadmap for on-premises deployment of the 3.7 generation. For organizations that can accept cloud inference and need multimodal GUI-agent capability, Claude Opus 4.8 and GPT-5.5 run at $5.00 per million input tokens, more than twelve times the Qwen3.7-Plus rate.

Frequently Asked Questions

Is Qwen3.7-Plus Open Source?

No. Qwen3.7-Plus is proprietary and available only through Alibaba Cloud Model Studio’s managed API. The weights cannot be downloaded or self-hosted. Some developers have speculated about open-weight releases based on Alibaba’s historical cadence, but the company has not announced a timeline, and third-party reports of a Q3 2026 open-weight variant are unconfirmed speculation.

How Does Qwen3.7-Plus Compare to Qwen3.7-Max?

Qwen3.7-Max is text-only, priced at $2.50 per million input tokens and $7.50 per million output tokens. Qwen3.7-Plus adds image and video input at $0.40 per million input and $1.60 per million output. On pure-text coding evaluations, Max holds a narrow edge; on any task involving screenshots, PDFs, UI navigation, or video, Plus is the applicable model and costs a fraction of Max per token. Both are API-only with no downloadable weights.

Can Qwen3.7-Plus Process Video?

Yes. The model accepts text, images, and video as input and returns text output only. Visual generation is handled by separate Alibaba model families. Video frames count as visual tokens and share the model’s one-million-token context budget, so large video payloads reduce available token headroom for accompanying text prompts in the same call.

What Is the preserve_thinking Parameter?

The preserve_thinking API parameter instructs Qwen3.7-Plus to retain its internal reasoning blocks across consecutive conversation turns. For multi-step agentic workflows, this prevents state decay: the model maintains its analytical chain across dozens of tool calls rather than recomputing context from scratch at each turn. Anthropic calls an equivalent feature “Extended Thinking” for its Claude models; OpenAI implements a similar mechanism as encrypted reasoning pass-back in GPT-5.5.

On Hugging Face, the Qwen3.6-35B-A3B weights remain available under Apache 2.0; Alibaba has released no equivalent for the 3.7 generation.