Research and Innovation

We will go all-in on AI in the intelligent era. We are committed to creating value for customers to achieve our own business success, while remaining driven by business value and market needs. This requires us to concentrate our innovation efforts on strategic businesses and ramp up efforts to make breakthroughs in basic theory application. In doing so, we will continue to sharpen the competitive edge of our products and solutions for the intelligent era, and serve our customers to the best of our ability.

Basic research: Concentrated innovation on strategic businesses and increased efforts for bigger breakthroughs in basic theory application

AI algorithms and application

• To address the high computational cost of multi- step sampling, we proposed a reinforcement- learning-guided sampling optimization algorithm. This algorithm balances generation efficiency and quality for multimodal diffusion models. Video and image generation is now five times faster, with accuracy loss below 1%.
• For better on-device LLM inference acceleration, we put forward a reinforcement-learning- augmented distillation technique. Thanks to a speculative decoding architecture that integrates both tree search and relaxed token acceptance, this new technique has doubled the inference throughput of Celia Auto-answer while maintaining the same level of accuracy.
• In long-tail scenarios, samples are typically sparse, while data collection suffers from poor efficiency. Furthermore, end-to-end generative modeling in high-dimensional perception-action spaces is a challenge in itself. To tackle these challenges, we have introduced a cloud-based World Engine and an on-vehicle World Action Model for vehicles. These have significantly boosted the intelligent driving capabilities of HUAWEI ADS 4, facilitating its deployment on over 1 million vehicles.

Communications technology and application

• Massive MIMO has huge demand for compute. However, by adopting the approach of iterative forward-backward optimization, we have surmounted this challenge. This approach takes us a long way towards discovering the representation of a high-precision matrix in extremely low-precision spaces, thus reducing the matrix computational complexity of transceivers by around 50%.
• We also developed an approximate operator design theory based on confidence to lower the computational complexity of digital pre-distortion (DPD) networks by approximately 20%.
• Wi-Fi signal interference and the high chip power consumption of 50G PON systems represent a major headache for the industry. In response, we have advanced a multi-level, self- adaptive interference cancellation architecture alongside a self-adaptive, dynamic bit-flipping algorithm. These have helped us lower system power consumption by around 10% without electromagnetic shielding.

Open innovation: Focus on market needs and business value for better competitiveness and greater value for customers

Computing

As compute is central to AI development, Huawei has worked diligently on our computing SuperPoD and cluster solutions in a bid to meet ever- increasing demand for compute. This has included:

• Releasing the roadmap for our future-oriented SuperPoD products, namely the Atlas 950 SuperPoD and Atlas 960 SuperPoD. These will have up to 8,192 and 15,488 Ascend neural network processing units (NPUs) respectively, and be ahead on all fronts, including key indicators such as NPU count, total computing power, memory capacity, and interconnect bandwidth.
• Building reliability into every layer of interconnect and working out a peer-to-peer architecture and unified protocol – UnifiedBus 1.0 that is centered on memory semantics – to ensure long range, high reliability, high bandwidth, and low latency. This also helps realize unified memory addressing, ultra-low, 100-nanosecond-level latency, and terabyte bandwidth, thereby making large-scale SuperPoDs a reality.
• Releasing UnifiedBus 2.0, an interconnect for SuperPoDs, in September 2025, and making the technical specifications public to drive both interconnect technology and the industry forward.

Wireless communications

To prepare for the age of mobile AI, we are embracing change through technology convergence and building a solid network foundation. With this in mind, last year, we:

• Broke through the bottleneck of high-dimensional channel pilot congestion in centimeter-wave T-MIMO by leveraging quasi- stationery propagation and scenario-aware fine- grained measurement of the spatial domain.This guarantees an approximately eightfold improvement in spectral efficiency.
• Used an integrated passive front-end radio frequency (RF) module for mmWave T-Cells and became the first company to realize sub-band frequency division multiple access (FDMA). This supercharges base stations, allowing them to easily handle heavy data traffic from a large number of concurrent users.
• Proposed an all-new approach to network- wide full-band synchronous turn-off for zero- energy-consumption timeslots and multi-level energy consumption configuration for refined scheduling. This cuts energy consumption throughout the entire process by nearly 50%.
• Continued to advance generative-AI-powered core networks using multi-agent technology, which was coupled with the centimeter- wave's integrated sensing and communications capabilities. This allows us to customize networks on demand for on-device AI and embodied AI, all while boosting wireless connectivity experience.

Optical networks

In optical networking, we:

• Made breakthroughs in demarcating, suppressing, and compensating for optical link impairments and a new forward error correction (FEC) dynamic decoding architecture, focusing on the light source mechanism, fiber channels, and algorithms. This enables highly- reliable and low-power optical interconnect between data centers.
• Broke ground on a number of key technologies such as laser communication in space, diffraction-limited optical antennas, and high- precision acquisition, tracking, and pointing (ATP). These advancements allow us to realize fast and high-precision inter-satellite tracking and acquisition, and support stable free-space laser communication.

Networking

In networking, we continue to enhance the performance of large-scale communications for SuperPoDs and clusters. To this end, we:

• Developed the industry's first Message-level signaling Congestion Control (MCC) algorithm, based on pathset information, to address demand for high throughput in large-scale cluster training and inference. This algorithm has helped us surmount the challenge of multi- path and per-packet congestion and improve communications throughput by 25%–50%.
• Created the industry's first parallel communications architecture for the data and control signaling planes. This delivers low latency for mixture of experts (MoE) inference for SuperPoDs. By reducing the synchronization and order preserving overhead, the architecture shortens collective communication completion time by 10%–30%.

Consumer business

In this domain, we remain at the forefront of technological innovation and experience, spanning from materials to architecture, terrestrial to satellite networks, and smart to intelligent application, as well as covering lighting and color. Specifically, over the past year, we have:

• Released the industry's first Switchable Dual Telephoto Camera. This camera perfectly combines a large sensor and dual telephoto lens to support dual optical zoom at 3.7x and 9.4x.
• Crafted a three-segment mortise-and-tenon Advanced Precision Hinge System with auxiliary cams to ensure our foldables deliver a seamless folding and unfolding experience.
• Introduced dual 3.5-mm ultra-thin fans and a copper-steel composite anti-gravity 3D vapor chamber to our foldable laptops, guaranteeing excellent cooling despite an unfolded thickness of just 7.3 mm.
• Enabled the terrestrial network to exceed 1 Gbps uplink peak rate and used an integrated chip architecture for both satellite and cellular communications for the satellite network, thereby guaranteeing seamless connectivity across ground and air environments.
• Provided a novel technology for emergency communication in network-blind spots. This has delivered a superior communications experience across terrestrial, satellite, and no-network environments.
• Achieved a quantum leap in our True-to-Color Camera. Supported by Multi-spectral Local White Balance technology, the camera supports pixel- wise white balance and works seamlessly with the Primary Camera, Ultra-Wide Angle Camera, and Telephoto Camera to deliver breathtaking color accuracy and ensure every hue is true to life in multiple-light-source scenarios.
• Employed AI to turbocharge computational aesthetics and graphics computing, delivering intelligent media experiences such as Custom Styles, Instant Movies in the Gallery, Immersive Light-field Visual Effects, and 3DGS Rendering Acceleration.
• Pushed the boundaries of AI through the deployment of on-device models, thus realizing more intelligent experiences such as call noise reduction, call summaries, Celia Brief, document summaries, and local AI searches, and making Celia even smarter and more capable.
• Completed the in-depth vertical integration of hardware, software, and chips to realize an over 35% performance improvement on our Mate 80 series phones with the HongMeng Kernel, BiSheng Compiler, and the OS Window System.

Foundational software

In this domain, we fully embrace AI to support critical foundational software capabilities for next-generation heterogeneous computing infrastructure, and to help Huawei provide a new option for the world with our solid AI computing foundation. Therefore, we have:

• Empowered the BiSheng Compiler to improve the performance of key Ascend operators by more than 20% through microarchitecture affinity compilation.
• Leveraged the Unified Memory Development Kit (UMDK) for communications acceleration, helping Ascend rise to the leading tier in terms of MoE Expert Parallelism inference performance.
• Achieved pioneering breakthroughs in technologies like intelligent awareness and scheduling, Optimized Container for multi-CPU clusters, and application-aware compilation acceleration. On this basis, our openEuler OS and BiSheng Compiler have helped Kunpeng secure leading performance in key scenarios.
• Broke ground on the pooling-based multi- write architecture for our GaussDB. When this is coupled with TaiShan 950 SuperPoD, we can enable smooth mainframe and midrange server migration for the core systems of financial institutions.

R&D engineering

In this domain, we have ramped up efforts to build efficient and trustworthy engineering capabilities. Over the last year, this has led us to:

• Develop a reliable, traceable, end-to- end verification engine for floating-point computations. This engine detects over 90% of silent data corruptions on NPUs, significantly increasing the likelihood of successful convergence for model training on ultra-large- scale computing clusters.
• Pioneer a new approach for enhancing the reliability of LLM-powered AI agents. Grounded on the Planning Domain Definition Language (PDDL), this highly adaptive approach is capable of qualitatively identifying risky task execution steps and intercepting 99.9% of the highest-risk operations.
• Reshape software engineering by ensuring that AI plays a bigger and more reliable role in request distribution, code generation, code review, unit testing, system testing, and open- source vulnerability remediation.