Alibaba's Qwen3.7-Plus adds vision and video — but closes the open-source door

The newest Qwen model is cheaper and more capable than its predecessor. It's also proprietary, which matters more than the benchmark numbers.

Written by Lena Armitage · Bureau Tech · Jun 3, 2026

The surprising part isn't the price

Alibaba's Qwen3.7-Plus costs $0.40 per million input tokens and $1.60 per million output tokens — genuinely cheap for a multimodal model at this capability tier. But the more consequential detail is what Alibaba gave up to get here: open weights.

Every major Qwen release up to this point shipped under Apache 2.0 or a comparable open-use license, letting enterprises download, audit, and self-host the models. Qwen3.7-Plus breaks that pattern. It's available exclusively through Alibaba Cloud's managed API. You can't run it on your own infrastructure.

For the many organizations — including, reportedly, Airbnb — that built workflows on open Qwen models precisely because they could control the deployment environment, that's a meaningful change.

What the model actually does

Qwen3.7-Plus is a large language model (LLM) that accepts text, images, video, and screenshots as inputs. Its predecessor, Qwen3.7-Max, handled text only. The new model is designed for agentic workflows: tasks where a model doesn't just answer a question but executes a sequence of actions — reading a codebase, interpreting a UI screenshot, running terminal commands — across many steps.

To support that, the model ships with a one-million-token context window and allocates up to 256,000 tokens for internal chain-of-thought reasoning — the step-by-step logic a model works through before producing an output.

The API exposes a parameter called `preserve_thinking`, which retains that reasoning state across conversational turns. Without something like this, long-running agent tasks tend to suffer from state decay: the model loses track of earlier logic and starts recomputing from scratch. Anthropic calls its equivalent "Extended Thinking"; OpenAI uses an encrypted reasoning pass-back mechanism. The underlying problem is the same across labs.

What the benchmarks show — and don't

On ScreenSpot Pro, which tests a model's ability to locate and interpret elements within interface screenshots, Qwen3.7-Plus scored 79.0. That's notably higher than GPT-5.4 at 67.4 and Claude Opus 4.6 at 49.5, per figures cited by VentureBeat.

On Terminal Bench 2.0-Terminus, which measures safe iterative terminal-level code execution, it scored 70.3, ahead of DeepSeek-V4-Pro Max (67.9) and Gemini 3.1 Pro (63.5).

Those are real numbers worth noting. But the source article is explicit that Qwen3.7-Plus still falls below leading U.S. proprietary models — including Claude Opus 4.6 and GPT-5.4 — on broader capability metrics. The ScreenSpot Pro result is a genuine outlier in its favor; it shouldn't be read as overall superiority.

The compliance question enterprises can't skip

Because Qwen3.7-Plus is API-only, all inference runs through Alibaba Cloud's international endpoints. The source documentation highlights a Singapore instance. That means any data sent to the model — code, screenshots, documents — leaves your environment.

For organizations subject to HIPAA, GDPR, or defense-sector data-residency rules, that routing requires explicit legal review. It's not a dealbreaker by default, but it's not a detail to paper over with a pricing comparison.

The managed API does remove the burden of provisioning GPU clusters to self-host. That's a real operational tradeoff. But it's a tradeoff, not a free lunch.

Where it fits

At $0.40 input / $1.60 output, Qwen3.7-Plus sits in a competitive band alongside MiniMax-M3 ($0.30/$1.20) and well below GPT-5.4 ($2.50/$15.00) or Claude Opus 4.8 ($5.00/$25.00). For high-frequency agent loops that process visual interfaces or large static codebases — and where data-residency rules permit cloud routing — the cost math is genuinely favorable.

For teams that need on-premises deployment, or that built on Qwen specifically because of its open licensing, this release doesn't serve them. That gap is worth naming clearly.

Key takeaways

Qwen3.7-Plus is multimodal (text, image, video, screenshots) where Qwen3.7-Max was text-only, at 60% lower cost.
It scores 79.0 on ScreenSpot Pro — a computer-vision interface benchmark — outpacing GPT-5.4 (67.4) and Claude Opus 4.6 (49.5), though it trails leading U.S. models on other agentic tasks.
The model is proprietary and API-only, ending Alibaba's run of Apache 2.0 open-weight releases; enterprises cannot self-host or air-gap it.
A 'preserve_thinking' parameter retains chain-of-thought state across multi-turn agent loops — Alibaba's version of a feature now standard across major labs.
Cached input pricing drops to $0.04 per million tokens for static context, making high-frequency agent loops more economical.

FAQ

What makes Qwen3.7-Plus different from Qwen3.7-Max?

Qwen3.7-Max is text-only. Qwen3.7-Plus adds image, video, and screenshot inputs, and is priced approximately 60% lower. The tradeoff is that Plus is proprietary and API-only, while earlier Qwen models were open-weight.

Can enterprises self-host Qwen3.7-Plus?

No. Unlike previous Qwen releases distributed under Apache 2.0 or similar licenses, Qwen3.7-Plus is available only through Alibaba Cloud's managed API. There are no downloadable weights.

What is 'preserve_thinking' and why does it matter for agents?

'preserve_thinking' is an API parameter that retains the model's internal chain-of-thought reasoning state across multiple conversational turns. In long-running agentic tasks, this prevents the model from losing its analytical thread mid-execution — a problem sometimes called state decay. Anthropic and OpenAI offer functionally equivalent features under different names.

How does Qwen3.7-Plus perform on benchmarks relative to competitors?

It scores well on interface-understanding tasks: 79.0 on ScreenSpot Pro, ahead of GPT-5.4 (67.4) and Claude Opus 4.6 (49.5). On terminal execution benchmarks it also leads several rivals. However, the source reporting notes it still trails leading U.S. proprietary models on broader capability evaluations.

What are the data-residency implications of using a cloud-only model?

All inference runs through Alibaba Cloud's external endpoints. Organizations subject to HIPAA, GDPR, or sovereign data requirements need to verify whether routing data through those endpoints — including a Singapore instance mentioned in developer documentation — complies with their specific obligations before deployment.