According to Pandaily, Huawei’s open-source agent platform openJiuwen added full-duplex multimodal interaction to WorkSwarm in an October 1 update. Pandaily said the change lets the assistant receive and deliver voice at the same time, while also creating a reported route to deploy full-duplex multimodal models on Ascend NPUs through Huawei Cloud ModelArts and the vLLM-Omni inference framework.
The update focuses on how a voice-and-video agent behaves when the conversation changes midstream. Pandaily said that when WorkSwarm identifies a new valid voice input, it halts its current spoken response and handles the new instruction. Text already produced stays in the task record. The account also says that voice-activity detection and a noise gate help WorkSwarm distinguish valid speech from interruptions that should not change the exchange.
Full-Duplex Interaction Reshapes the WorkSwarm Conversation Loop
That sequence gives the October update a defined interaction model. Listening and speaking occur together, rather than requiring a user to wait for an answer to finish before supplying the next instruction. When a valid new request arrives, the spoken response stops, but the previously generated text is retained in the record. Pandaily attributed both that behavior and the use of voice-activity detection and a noise gate to the WorkSwarm update. The details describe interruption handling, not a measurement of response time, recognition quality, or reliability.
Pandaily also described a two-track task arrangement behind WorkSwarm. A real-time model handles camera input, listening, and quick replies. Separately, a Core Agent decomposes tasks, invokes tools, and keeps longer work moving in the background. This division matters for the reported task controls: users can reorder queued work and stop work that is queued or already running. Ending or redirecting a spoken exchange does not, by itself, cancel tool tasks that continue under the Core Agent.
The distinction separates the real-time conversational layer from longer-running task execution without identifying the exact multimodal model or model version used by WorkSwarm. Pandaily did not provide an Ascend chip model, latency figures, throughput, pricing, user totals, or customer deployments. It also did not establish that the feature is commercially deployed at scale. WorkSwarm’s new behavior therefore has to be read as a product update with defined controls, rather than as an independently measured performance claim.
For WorkSwarm, the documented point is narrow: an interrupted conversation does not automatically terminate background tool activity, and users retain controls over the queue. The October update does not supply a benchmark or customer case study for that design. The agent-workflow theme also appears in EastFrontier’s recent Huawei coverage.
vLLM-Omni Connects the Update to a Versioned Ascend Software Stack
Pandaily said that full-duplex multimodal models can be deployed on Ascend NPUs through ModelArts using vLLM-Omni. The framework itself is documented by the vLLM-Omni maintainers’ GitHub repository as an Apache-2.0-licensed project for inference and serving of omni-modality models. Its description covers text, image, audio, video, and action data, with outputs that can include text, multimodal material, and actions. Those are framework documentation claims, rather than evidence about WorkSwarm’s exact model configuration.
The repository lists real-time full-duplex serving with streaming audio input and output as a supported feature, alongside a realtime duplex API. Its September 2026 notes say version 0.30.0 added an engine-session architecture for serving MiniCPM-o 4.5 and AURA in full duplex. The notes also list cross-stage KV and multimodal-payload transfer, and describe real-time MiniMax H3 Turbo support across NVIDIA Blackwell and Ascend 950. None of those named models or hardware references identifies what WorkSwarm uses, and they should not be read as a WorkSwarm benchmark.
The project’s release history establishes that its full-duplex and NPU work predates the October WorkSwarm announcement. Version 0.28.0 added what the maintainers describe as production-ready MiniMax H3 serving on GPU and NPU, and enhanced real-time full-duplex serving for the MiniCPM-o series. Version 0.26.0 introduced an experimental full-duplex real-time runtime for MiniCPM-o 4.5. The documentation also lists model categories that include Qwen3-Omni, MiniCPM-o 4.5, Qwen3-TTS, CosyVoice3, diffusion models, and robot-policy models. That breadth explains the framework’s stated scope but does not reveal the model behind WorkSwarm’s update.
The vLLM-Omni NPU installation documentation adds practical limits to the reported Ascend pathway. It says NPU support comes through the community-maintained vLLM Ascend Plugin, or vllm-ascend, for running vLLM on NPU. The page specifies Linux and Python 3.12, says vLLM-Omni is not natively supported on Windows, and recommends prebuilt vLLM-Ascend Docker images. It lists A2, A3, A5, and 310P variants, with 310P marked experimental, and gives an example verified on four NPUs using the v0.30.0 image line.
Those server-side conditions do not contradict the official WorkSwarm page, which lists HarmonyOS, Windows, and Mac as product platforms. The product page says WorkSwarm is available in the HarmonyOS PC App Market and displays installation through the workswarm and workswarm-init pip commands. It also lists task planning, self-evolution, context slimming, portable memory, browser-control optimization, and Xiaoyi and Feishu integration as capabilities. These are official product-page descriptions, not independently tested results or confirmation that every product channel uses the NPU deployment path.
Huawei’s Ascend Claims Frame the Infrastructure Context
Huawei’s September 17 HUAWEI CONNECT keynote provides company-reported context for why a ModelArts and Ascend pathway matters to the company’s software stack. In the Huawei keynote, the company said Ascend supports more than 90 leading third-party open-source projects, naming PyTorch, Triton, vLLM, and veRL, and said Ascend is officially supported as a PyTorch accelerator backend. Huawei presented those statements as ecosystem claims, not independent certification of WorkSwarm or its full-duplex behavior.
Huawei also said external CANN developers made up 61% of all CANN developers, that the platform had more than 5,200 monthly active developers, and that more than 40 models had been natively pre-trained on Ascend and CANN. The keynote did not tie those figures to WorkSwarm, vLLM-Omni, or the October update. Likewise, Huawei said more than 1,000 Atlas 900 A3 SuperPoDs had been deployed and that Atlas 950 SuperPoD was seeing large-scale commercial use. Those are Huawei figures, not evidence of a WorkSwarm customer deployment.
The confirmed WorkSwarm details remain specific. Pandaily identified ModelArts and vLLM-Omni as a deployment route for full-duplex multimodal models on unspecified Ascend NPUs. The framework documentation shows a versioned, dependency-sensitive NPU route, while WorkSwarm’s own page describes its platforms and feature labels. Related EastFrontier archive coverage provides additional Ascend context. Together, those records document an announced interaction update and a stated deployment pathway, without establishing exact hardware compatibility, user adoption, or measured application performance.