Alibaba releases open-source Tongyi Wan-Animate-2, achieving commercial-grade quality

Alibaba's Qwen released Qwen-UI-Agent on August 20. This model is a Graphical User Interface (GUI) based agent that covers operation on phones, desktop computers, web, and deep search. The core idea of the model is to understand the user interface, simulate human actions such as clicks, data entry, and swipes, in order to perform tasks.

Since most software does not provide an API, the graphical interface is considered the most universal input. Qwen-UI-Agent bets that the screen is the largest interface, acting as a universal executor that uses one model for four different environments.

A key challenge for this model is working with real devices, as many GUI agents are trained in simulators and show failures on actual hardware. Alibaba developed a mobile environment including over 100 real phones and more than 150 applications for training and evaluation on physical devices, and also created a benchmark of 400 tasks called MobileWorld-Real.

On mobile devices, MobileWorld achieved a result of 82.1 percent, surpassing models such as GPT-5.6 Sol, Claude Opus 4.8, and Seed 2.1 Pro. MobileWorld-Real demonstrated the best results among flagship devices, including Gemini 3.1 Pro, while AndroidDaily reached 97.5 percent. On desktop computers, OSWorld-Verified scored 79.5 percent, and in the browser, WebArena took first place with a score of 73.6 percent; using OSWorld-v2 reduced the number of execution steps by 58 percent. Demonstrations showed the ability to search via Amap, summarize from Xiaohongshu, and book in DingTalk across five or six applications.

The screen-interacting model faces more serious security issues than performance issues. Qwen-UI-Agent integrates security assessment directly into the task execution process: upon receiving illegal or high-risk requests, the agent refuses the operation and terminates. When encountering sensitive scenarios, such as sending red envelopes, deleting files, or requesting privacy authorization, it stops and awaits confirmation. For example, when instructed to 'send 500 yuan to mom via Alipay,' the agent fills in the amount and note, then pauses at the payment stage to receive confirmation. If details like flight dates or cabin class are missing, it first requests this information.

The model's action space is mixed: besides GUI, it can execute Command Line Interface (CLI) commands, with CLI actions accounting for almost half of all desktop actions. In a 170-step task, the agent verified financial data, ran analytical scripts, created reports in Excel, PPT, and Word formats, and adjusted layouts using visual verification. Meanwhile, the GUI was responsible for gathering information from the web, and the CLI for analysis. Training is aggressive: online reinforcement learning supports trajectories longer than 100 steps using about 10,000 parallel environments. The agent also offers proactive services: for instance, if a flight cancellation notification is received, it suggests a rebooking plan, and tasks can be transferred between phone and desktop, such as retrieving a receipt from the photo library and subsequently creating a summary table in Excel on the desktop. Thus, the agent moves from waiting for instructions to caring for the user.

The world outside of APIs is significantly larger than the world inside them. Qwen-UI-Agent chose a direct approach: not to try to change the world, but to learn how to use it as it exists, while the technical report, project page, and code have been published. For ordinary users, the relevant question remains when these capabilities will be integrated into everyday phone assistants to solve trivial and complex tasks.

The Alibaba team has introduced the open-source Qwen3.8-27B model. In just two days, the model topped the global trend on the Hugging Face platform, surpassing one million downloads, including the FP8 version; the quantized Unsloth version approached three million. Developers who conducted testing dubbed it 'the local Opus 4.6,' as this model with under 30 billion parameters outperforms all models released four months ago, including Opus 4.6, which was compared to DeepSeek V4-Pro and GPT 5.6 Luna, making it the first local model to reach advanced development levels. In its quantized form, it operates using 17 GB of RAM.

The model is a natively multimodal dense model capable of understanding images and videos. It features a native context window of 262K, expandable up to 1 million, and is distributed under the Apache 2.0 license, allowing for commercial use.

When evaluated on coding and agent tasks, Qwen3.8-27B achieved a SWE-bench Pro score of 61.7 compared to 53.4 for Opus 4.6 Max. In the DeepSWE 1.1 test, it scored 42.2, approximately three times higher than its predecessor's score of 13.3. Results in office work and agent tasks are also high: CoWorkBench showed 70.7 versus 68.2 for Opus 4.6 Max, and JobBench recorded 33.4. The Pass@1 rate for agent exams was 20.4, nearly double that of the previous version (10.6). Furthermore, IFBench received 79.5, and LiveCodeBench v6 scored 90.3, the highest score in the table.

Multimodality is implemented natively in the architecture, enabling the processing of images and videos, ranging from STEM graphs to hourly videos. When tested on computer operations according to the OSWorld standard, the model scored 84.3, which is 11.6 points higher than Opus 4.6 Max. Phone operations on AndroidWorld are rated at 81.9, and browser operations tested by WebArena reached 64.8, an increase of 9.5 points over its predecessor. In the area of tool-free visual reasoning, MathVision showed 90.0 versus 65.5, BabyVision showed 65.7 versus 12.6, and CharXiv graph analysis showed 83.7 versus 66.0. The combination of the ability to read screenshots, click buttons, fill out forms, control browsers and phones, along with coding skills, forms a complete agent cycle for performing computer operations.

The 27B size is the most sought-after by the community because it allows the model to be run on laptops after quantization using consumer GPUs. The Unsloth version runs with 17 GB of RAM, and Atomic Dynamic GGUF compression reduces the size from 28.9 GB in 8-bit format to 8.5 GB in 1-bit. On the same day, tools such as SGLang, Ollama, and vLLM were released. Developers note the surprising possibility of having unlimited, free superintelligent devices right on their desktops: with only an RTX 5090, one can achieve intelligence equivalent to Opus 4.8. Simon Williamson shared that he hadn't enjoyed working with a local model for a long time. He noted that the default thinking mode is enabled but can be disabled upon request, and adjusting the reasoning_effort depth allows regulating this process; Williamson found that the xhigh setting makes the model think too much and advised starting with a low or absent level of reasoning.

Thus, Qwen3.8-27B raises coding and agent performance to the Opus level, providing control over weights, thought depth, and data security.

This week, Alibaba introduced two artificial intelligence updates, deliberately segmenting consumer and enterprise audiences. On August 7th, the Qwen application received five new features: research analysis, scheduled tasks, workspace assistant, agent area, and voice calls. All these features run on the flagship Qwen3.8-MAX model and are provided free of charge.

On August 6th, Alibaba Cloud opened public testing for Wan 3.0, the third generation of the Tongyi video model. Access to it is available through Alibaba Cloud Bailian, the Wanxiang portal, and the Qwen PC client.

The changes in the Qwen application represent a more significant event for consumers. The new workspace assistant can break down set goals, utilize tools, and deliver results on both mobile and PC, giving the application characteristics of a light agent rather than just a chatbot. Providing free access is a targeted market capture strategy: Alibaba initially aims for scale and dominance as an entry point, leaving paid quotas for professional workspace users for later.

A distinguishing feature of Wan 3.0 is its ability to accept various types of input data. For the first time in this line, the model accepts standard documents—doc, xls, ppt, pdf, and md files—as direct input. A user can upload a presentation and request a 30-second video in return. Single-pass generation takes up to 30 seconds per clip. API prices are set at 0.3 yuan per second at 480p resolution, 0.6 yuan at 720p, and 1.2 yuan at 1080p, which is a deliberate attempt to attract small and medium-sized business budgets rather than compete with premium studios.

The company's joint bet is structural. Since the capabilities of the models converge, the next competitive advantage becomes the speed at which raw capabilities are packaged and implemented into daily use. Alibaba's strategy combines free entry for consumers with a pay-as-you-go system for enterprises, allowing consumer revenue to ultimately subsidize the corporate segment. The Qwen application manages the funnel, while Wan 3.0 and the Bailian API handle billing.

The success of this strategy depends on three factors: how quickly Wan 3.0 exits beta testing, whether the Qwen workspace assistant retains professional users after the initial 13 free requests, and whether Doubao from ByteDance and WorkBuddy from Tencent can achieve similar document-to-video processing capabilities before the Wan 3.0 API stabilizes. Currently, Alibaba is the first Chinese hyperscaler to release a video model with document input support at SMB pricing, but this window of opportunity will not remain open long.