Huawei Cloud releases Harmony Coding LLM tailored for ArkTS

Huawei Cloud CEO Zhou Yu Feng announced at the Huawei Connect 2026 event plans for the commercial launch of a cloud service based on the Ascend 950 Lingqu cluster. The launch in China is scheduled for September 30, and access for the global market will open on November 30.

These dates define the timeline for Ascend 950 compute SKUs and do not merely describe the SuperPoD architecture. Customers will be able to reserve a cluster connected via UnifiedBus consisting of 1024 cards as a managed cloud offering according to the published schedule.

Huawei stated that the Ascend 950 system utilizes its UnifiedBus interconnect, marketed as Lingqu, to assemble a cluster of 1024 accelerators. This cluster is capable of providing up to 1 EFLOPS of computation in FP8 format and 2 EFLOPS in FP4 format, and also features 256 TB of unified memory available globally.

The design is engineered to support end-to-end training of large models without the need for additional cluster partitioning or adaptation, maintaining thousand-card job execution within a single memory address space.

The commercial window for cloud services fits into a broader Ascend development roadmap presented at the same event. Chairman Wang Tao noted that over 1000 Ascend supernodes have already been deployed, and Ascend 950 supernodes have entered a phase of scalable commercial use.

Furthermore, the Ascend 960DT is planned for release in the first quarter of 2027, and the Ascend 960PR in the third quarter, with a stated release pace of approximately one major Ascend generation annually, including long-term stages 970 and 980.

For potential buyers, the key information is the specific calendar with the launch first in China and then globally, as well as the cluster specification: 1024 cards, EFLOPS class throughput (FP8/FP4), and 256 TB of unified memory via UnifiedBus. Huawei Cloud positions Ascend 950 Lingqu as a ready-to-use AI cluster service, rather than a theoretical architecture, offering internal availability at the end of September and international availability two months later for teams needing supernode-scale training without building the infrastructure themselves.

ByteDance announced that the updated version of Doubao-Seed-2.1-pro, release 0915, is fully available via the Volcano Ark API. Concurrently, Doubao Work has been updated, and the TRAE system is now connected. The company emphasizes the implementation of multimodal encoding in production processes, enhancing agent reliability, and reducing token costs for processing images and videos, which, according to the company, is more than 30% lower than the previous generation.

Multimodal encoding is presented as a key feature of the product. ByteDance asserts that corporate tasks are often found not in formal documents but in screen recordings, UI sketches, and operator habits. In an official case study, Doubao-Seed-2.1-pro 0915 was able to analyze undocumented Java ERP, consisting of approximately 280,000 lines, using a recording and several sketches. The model successfully derived the architecture and created a working mobile frontend, shifting the focus from humans writing machine-readable specifications to models reading human visual expressions.

Another demonstration example used the open-source game repository Luanti, which included about 387,000 lines of code and 1,000 historical issues, with official fixes hidden. The model organized the work of several sub-agents over almost 36 hours to perform root cause analysis and correct files among themselves; approximately 83% of the completed tasks met a standard acceptable to engineers.

Separately, the model restored an interactive three-dimensional scene of a seasonal courtyard based on four design images, including structure, water, vegetation, and camera effects. Volcengine also highlights agent reliability and cost-effectiveness. Previously, Doubao-Seed-2.1-pro held leading positions in several benchmarks such as Terminal Bench 2.1, SWE-Pro, SciCode, OSWorld, MobileWorld, and MMMU-Pro. Update 0915 adds stronger evidence tracking, source authority checking, and verification for extensive research reports, along with improved document and engineering drawing parsing. The increased efficiency of tokens for image and video processing is positioned for use in high-frequency corporate APIs. This release is an update to the model and product on the Volcengine platform, not a funding announcement, and external teams can still reproduce the ERP and Luanti merger metrics on their own stacks.