Huawei introduces TTSE, an agent framework that separates environmental facts and tips

Huawei has introduced its AI assistant as a standalone application in the AppGallery for HarmonyOS PC and officially released its agent for handling office tasks, according to IT Home on November 6. This desktop application is version 11.7.9.300.

Previously, this assistant was integrated directly into the computer's operating system. Now, as a separate application, it provides an update distribution channel outside of the system package. IT Home notes that the Work function underwent internal testing back in September, and the current release moves this office assistant beyond the previously described testing period.

Users can now assign tasks such as preparing presentations, analyzing data, and compiling weekly reports using just one command. According to the report, the agent breaks down the assigned goal into sequential steps, selects the necessary tools, and performs the work, generating finished documents for Word, PowerPoint, and Excel instead of just providing a conversational response.

On the Chinese product page, Huawei describes two operating modes: cloud execution on PC and local execution on the computer. Users can choose the appropriate environment depending on the software and files requiring processing. The company emphasizes that progress remains visible, allowing users to monitor important checkpoints rather than leaving all stages to the agent's discretion.

The service also supports connection to various office platforms, including DingTalk and Feishu. Huawei states that over 200 expert teams and specialists are available within the service. These statements are company claims and do not represent an independently assessed measure of task completion quality.

The connected HarmonyOS smartphone can remotely send tasks to the computer, track the progress, and receive completed results. Huawei indicates that tasks can continue even after the user leaves the PC, and one phone can be linked to multiple computers. Actual execution depends on the availability of the selected device, files, and environment.

IT Home reports that Work operates on a membership model, where task consumption is measured in AI points; more complex work requires more points. The report does not provide current subscription prices. The new development is the release of the autonomous PC distribution and the official office agent, which differs from Pandaily's November 5 report on the appearance of openPangu-2.1-Pro within the service. Huawei has not yet confirmed an official English name for this Work product intended for the Chinese market.

Huawei's AI agent, Xiaoyi Work, has become the first product to integrate openPangu-2.1-Pro—a new version of the company's open Pangu model family. This was revealed following testing conducted by the publication IT Home on November 5. The application states that this model is intended to improve tasks within the HarmonyOS operating system.

The appearance of this model occurred several days after Huawei Cloud announced on November 1 that it would cease sales of the openPangu-2.0-Flash and openPangu-2.0-Pro models on November 6, decommissioning them by midnight on November 30. At the time, Huawei reported plans for new openPangu models that would be rolled out gradually.

Details about the openPangu-2.1-Pro model have not yet been published by Huawei. There is no public information regarding the number of parameters, active parameters, context length, or training data used. The model has not been officially released, and IT Home concluded that it is a very new development available exclusively within Xiaoyi Work. For comparison, its predecessor, openPangu-2.0-Pro, was introduced earlier this year with a total of 505 billion parameters, 180 billion sparsely activated parameters, and a 512K context window, trained entirely on Ascend NPU.

Xiaoyi Work is positioned as an AI assistant for performing office, coding, and creative tasks on HarmonyOS devices, including phones, tablets, and PCs, and also features a Windows client. The user sets a goal for the agent, and it breaks down the task into stages, calls necessary tools, and performs the work, providing the user with visible progress and key checkpoints for verification. Tasks can be executed either on a cloud PC or a local computer, while the phone can monitor the process running on a connected PC.

According to a review from InfoQ China, this product was developed jointly by Huawei's Xiaoyi team and the openJiuwen community of open agents. The service is in beta and is paid. New users receive 1000 free AI credits for 31 days; after that, the basic plan costs 49 yuan per month with automatic renewal, and the standard plan costs 199 yuan per month. Supported scenarios include market analysis, document processing, slide creation, data analysis, content generation, and code development.

The first openPangu-2.1 model appeared within the agent product, rather than on the cloud model marketplace, with its stated focus being agent tasks on HarmonyOS devices. Huawei has not specified when openPangu-2.1-Pro will be officially released or if it will be available through Huawei Cloud MaaS.

openJiuwen, an open-source framework supported by Huawei, has released OmniMem-Lens—a multimodal memory feature in version 0.2.0 of its AgentMemory module. This feature allows AI agents to remember long videos based on actual events rather than fixed time intervals, according to QbitAI on November 4.

The team claims that the standard approach of splitting long videos into one-minute segments and storing each separately compromises agent memory in two ways. When a continuous scene is split across windows, the agent only receives half the evidence, which contributes to hallucinations. Furthermore, a few seconds of key action mixed with minutes of irrelevant material can be lost in the noise, reducing search accuracy.

Most event segmentation methods also rely on only one type of data, causing them to miss topic changes if the image remains static, or scene changes if no one is speaking. OmniMem-Lens tracks both streams: the visual track notes scene changes, while the audio track uses speech recognition to determine topic shifts by speakers, even if the screen does not change. Cross-referencing these two sources allows for cuts precisely at the boundaries of real events.

In one example, showing a two-hour product review where the same architectural diagram was displayed for 40 minutes while the speaker covered three different topics, the audio track separated them into three memories, enabling an answer to a subsequent question about load testing metrics.

Memory is stored in three levels: L0 retains the original video, L1 contains event-level memories, and L2 holds more detailed segment-level memories. Levels L1 and L2 are fully converted into text and vectors, allowing them to be integrated into existing augmented retrieval generation stacks and vector databases with minimal changes. The main data organization work is performed offline. During a query, the system matches the query against L1 and L2, extracting precise frames from L0 only when visual detail verification is required.

When testing OmniMem-Lens alongside Qwen3-VL-8B against the direct response of Qwen3-VL-8B from the video, information extraction accuracy increased by 8% on Video-MME-L and by 6.4% on LVBench. The largest gain, 45.7%, was achieved on EgoLife-QA, a first-person dataset recording a full day of daily life where fixed slicing often fails. On LVBench, online response time decreased from 34.5 seconds to 1.72 seconds, approximately 20 times faster, and the team estimates a maximum latency reduction of 97%.

The project's source code is available on GitCode with a mirror on GitHub under the Apache-2.0 license. The team reports that openJiuwen WorkSwarm will soon begin supporting AgentMemory, which will allow remembering and recalling videos exchanged in everyday conversations during subsequent discussions.