Alibaba opens Qwen3.7 — Plus multimodal agent APIs on Bailian (Model Studio)

Alibaba’s Qwen team has deployed Qwen3.7 — Plus to its Bailian platform (branded internationally as Model Studio) and opened API access for external developers. The release makes the model’s multimodal and agentic capabilities available for integration, enabling builders to embed planning, tool use and iterative workflows into applications rather than relying solely on single — turn conversations. This matters because it gives developers direct access to an agent — capable large model through a hosted console and programmatic endpoints.

Qwen3.7 — Plus follows the earlier unveiling of the Qwen3.7 generation in May; the Plus variant was made available in preview on the platform so builders could test the new multimodal and agent — style features. The preview phase was presented as a way to let integrators experiment with the model’s expanded skill set before broader production rollout.

In public evaluations the preview produced measurable vision results: Qwen3.7 — Plus ranked #16 overall in Vision Arena, which placed Alibaba as the #5 lab on that LM Arena public leaderboard. Those rankings were derived from blind matchups where users vote on image — understanding answers; the Vision Arena placement was highlighted as relevant to tasks such as OCR, chart reading and per-frame video analysis.

Alibaba frames the release around agent — style capabilities rather than only conversational responses. The company lists five core abilities targeted for Qwen3.7 — Plus: deep reasoning, self-programming, tool invocation, verification and testing, and autonomous iteration. Together these capabilities are intended to move the model from answering single prompts toward carrying out multi — step workflows in which it can plan actions, call tools, evaluate outcomes and repeat until a goal is achieved.

Bailian’s hosting environment adds two platform — level components designed for builders: an agentic reinforcement learning mechanism that uses real-world execution feedback to refine model behavior over time, and built — in safety guardrails that constrain autonomous tool usage. Those platform features are intended to make long-running, automated tasks more tractable while keeping external calls and file edits inside preset operational limits.

For engineers and integrators the agentic feature set is described in concrete terms: self-programming means the model can generate and revise its own code; tool invocation maps to calling external functions or APIs via the platform; verification and testing indicate the model can run outputs and check results; and autonomous iteration describes looping behavior that continues until the task completes. Bailian exposes these mechanics via API so external developers can integrate and experiment with the model’s agentic workflows.