Alibaba’s Tongyi Qianwen team released Qwen3.7-Plus on June 6, a multimodal model built for computer-use agent workloads. The model reads screens, clicks buttons, writes code, and operates terminal commands within a single agent loop, according to WinBuzzer.
Alibaba calls it a “multimodal interactive hybrid agent.” The label describes what the model does: it perceives screen state, plans actions, executes them through GUI clicks or CLI commands, and checks results. Previous computer-use models from Anthropic (Claude, October 2024), OpenAI (Operator, January 2025), and Microsoft Research (Fara1.5, May 2026) focused primarily on browser automation. Qwen3.7-Plus extends that scope to desktop applications, cloud consoles, and terminal sessions.
Benchmark Performance
The numbers favor Qwen3.7-Plus on screen-grounding tasks. It scored 79.0 on ScreenSpot Pro and leads on AndroidWorld, placing ahead of GPT-5.4, Claude Opus 4.6 Max, and Gemini 3.1 Pro on GUI-oriented benchmarks, according to The Decoder. It also tops Terminal-Bench for terminal-task completion and long-horizon task planning scores.
On harder scientific reasoning tasks like MedXpertQA-MM, the model falls behind Gemini 3.1 Pro and GPT-5.4. The Decoder describes text-side performance as “on par with max-tier models, without beating them across the board.”
The Demos
Alibaba published three demonstrations. A hybrid agent using Qwen3.7-Plus built an English vocabulary learning app over eleven hours, producing more than 10,000 lines of code across more than 1,000 agent calls, according to WinBuzzer. A second demo recreated the native macOS Stocks app: the agent parsed the UI, generated SwiftUI code, connected an API for real-time data, compiled the result, and ran ten functional tests. A third showed Qwen for Chrome entering agent mode to purchase and configure a cloud server instance.
These remain vendor demonstrations, not independent evaluations. Long-running agent workflows compound small errors across hundreds of steps, and production reliability is unproven.
Pricing and Compatibility
Qwen3.7-Plus costs $0.40 per million input tokens and $2.40 per million output tokens, according to both WinBuzzer and The Decoder. That is roughly six times cheaper on input and three times cheaper on output than Alibaba’s own Qwen3.7-Max ($2.50/$7.50), and well below the list prices of GPT-5.4 or Claude Opus 4.6.
The model supports the Anthropic API protocol and works directly with Claude Code, OpenClaw, and Alibaba’s Qwen Code. A preserve_thinking API feature retains reasoning content from earlier conversation turns, which the Qwen team recommends for agentic tasks.
Unlike its text-based sibling Qwen3.7-Max, Qwen3.7-Plus is a proprietary offering with no open weights. It is available through Alibaba Cloud Model Studio.
The Computer-Use Race
Alibaba is now the first major Chinese AI lab to ship a dedicated computer-use agent model. The competitive field has grown rapidly: Anthropic introduced computer use for Claude in October 2024, OpenAI launched Operator for browser actions in January 2025, and Microsoft Research released the Fara1.5 model family in May 2026 with browser agents in 4B, 9B, and 27B parameter sizes, as WinBuzzer documents.
Qwen3.7-Plus claims broader scope than any of those by combining screen, terminal, coding, and cloud-console operation in one model. Whether enterprises outside Alibaba’s existing cloud customer base adopt a proprietary Chinese model for desktop automation depends on factors beyond benchmark scores: data residency requirements, trust in audit trails, and whether the agent can reliably handle interface changes, permissions failures, and error recovery over sustained production runs.