Keep Company introduced its proprietary vertical sports model, Keepace.ai, at the WAIC 2026 event. This model prioritizes adherence to safety boundaries over general functionality. It was developed based on 14 billion records of real exercises and incorporates over 700 metrics in a panoramic feature graph.
The presentation of Keepace.ai marked the first public demonstration of the company's core artificial intelligence capabilities since announcing the 'All in AI' strategy in early 2025. The model addresses a critical issue in general AI systems for fitness applications: while universal models strive for maximum utility, they lack the necessary safety constraints, scientific rigor, and personalization detail required for proper physical load consultation, as incorrect advice can lead to physical harm.
The vertical approach was motivated by three main aspects. Firstly, safety boundaries: an error in a general chatbot requires only user verification, whereas the same error in exercise recommendations provided to a human can cause irreversible damage. The Keepace.ai architecture treats safety as a hard constraint embedded in the training and evaluation system, rather than merely an optimization goal. Secondly, filtering pseudoscientific claims: the fitness sector is one of the most marketing-saturated environments on the internet, where myths are widespread, such as the idea that the more you sweat, the more fat you burn. General models absorb and reproduce such misconceptions from training data, whereas Keepace.ai uses authoritative source verification as a strict evaluation criterion. Thirdly, personalization detail: transforming raw sensor data into an accurate model of a user's physical capabilities requires specialized training unavailable to general models.
Currently, Keepace.ai supports three use cases: generating workout plans, answering exercise knowledge questions, and interpreting exercise data. For instance, when requested to perform back exercises with dumbbells, accompanied by a message about recent back discomfort, the model recognizes the risk and suggests a single-arm row instead of bent-over rows or deadlifts to reduce strain on the spine. In Q&A mode, if asked about running pace for weight loss, the model first establishes that calorie deficit is the fundamental principle, and then refutes the common misconception that the body burns sugar before fat, explaining that all three energy systems work simultaneously with proportions depending on intensity. Every recommendation is supported by a link to authoritative sources verified by sports science consultants.
A significant advantage supporting Keepace.ai is the extensive dataset: 14 billion real exercise records collected over a decade, structured into a panoramic feature graph including 17 label categories and over 700 metrics. This structured long-term data represents a competitive advantage that cannot be acquired or replicated solely through investment in computing power. In 2025, Keep achieved its first annual profitability, showing adjusted net profit of 25.22 million RMB, overcoming the previous year's loss of 470 million RMB, while gross margin increased to 52.2%. The company views AI investments as cost replacements rather than additions, using AI to substitute traditional content creation costs while simultaneously enhancing service quality and enabling potential revenue growth tenfold. According to Keep's thesis, vertical AI models outperform general models where domain limitations render the latter unsafe. At its initial stage, Keepace.ai prioritizes safety over personalization, with the algorithm lead emphasizing that understanding sports means knowing what not to train and explaining why.
The Chengdu High-Tech Zone has established itself as a leading industrial cluster for space computing in China. At the APEC Digital Week in 2026, a comprehensive ecosystem was presented, covering satellite design, rocket manufacturing, ground terminals, and satellite applications.
This zone hosts 102 key commercial aerospace enterprises, which reached an industry volume of 13 billion yuan in 2025. A local industry leader successfully launched the world's first orbital artificial intelligence computing network using one rocket with 12 satellites. This created the first closed loop from AI computing in space to terrestrial application.
The core idea of space computing is to place data processing, storage, and AI analysis capabilities on low-Earth orbit satellites. This transforms them from simple data relay stations into intelligent computing nodes. Space provides conditions of near-perfect vacuum, extremely low temperatures, and virtually unlimited solar energy, allowing for a Power Usage Effectiveness (PUE) close to the theoretical ideal of 1.0, unlike ground data centers that face limitations in power consumption and cooling costs.
It is projected that the commercial potential will reach trillions of yuan, as the demand for AI inference requires computational power that ground infrastructure cannot economically provide at scale. Engineering challenges remain significant: standard GPU chips are susceptible to cosmic radiation, leading to rapid device aging, and radiation-hardened production is expensive. In response, the industry is exploring photonic computing technology, which is inherently immune to radiation, requires no charge, and is energy-efficient. Solar panel manufacturers are developing perovskite-silicon tandem cells optimized for the extreme thermal cycling of space. High launch costs and orbital laser networking remain major hurdles to mass adoption. Chengdu has already secured the base for final rocket assembly and engine manufacturing to reduce launch costs from the outset.
Chengdu's industrial strategy is based on three pillars: strong technologies through a space computing innovation consortium focusing on computing architecture, inter-satellite communication, and thermal management in space, as well as space-class energy technologies; smart operations, including a space computing standards alliance aimed at developing national and international standards by 2027; and high launch capacity thanks to its rocket manufacturing base. The zone manages the largest fund cluster in Central-Western China, comprising 182 funds totaling over 350 billion yuan, with 158.1 billion yuan in direct funding allocated in 2025 alone. A specialized aerospace industry fund of 2 billion yuan and a venture capital fund for space computing are also under development.
The future city of Chengdu has attracted 200 technology enterprises with investments totaling 130 billion yuan and claims the capability for full satellite delivery. As the demand for AI inference grows and ground limitations become critical, space computing is viewed not as speculation but as a necessary complementary element. Chengdu positions itself as the primary ecosystem for this emerging global industry.
MindLab has released Macaron-V1, a model fine-tuned using the Mixture-of-LoRA method. This development is built upon the foundation of Zhipu AI GLM 5.2 and demonstrates how targeted reinforcement learning (RL) fine-tuning can significantly boost the capabilities of an existing model with minimal computational overhead.
The flagship Venti version, boasting 748 billion parameters and supporting a million-token native context, utilizes four independent LoRA expert adapters with 1 billion parameters each. These experts cover chat, agents, coding, and user interface domains. Each expert was trained in an isolated adapter space using the proprietary Mixture-of-LoRA technique, which prevents the interference between capabilities often observed during multi-task fine-tuning.
The MoL method represents a significant departure from traditional fine-tuning. Standard multi-task fine-tuning suffers from the oscillation effect, where improving one function leads to the degradation of another. MindLab's solution places numerous independent LoRA experts atop a frozen base model, with each optimized in its own isolated adapter space. A router mechanism selects the appropriate expert for each query, and new capabilities can be added simply by connecting another LoRA expert without retraining the existing ones.
The technical innovations in Macaron-V1 extend beyond MoL. The model incorporates LongStraw technology, which extends GLM 5.2's native context to two million tokens by reusing common prompt segments as a persistent state, substantially reducing the overhead associated with long sequences. Furthermore, there is a lightweight Tall version with 35 billion parameters designed for local deployment based on Qwen 3.6.
To support the system, three components were jointly trained: GenUI, responsible for displaying interactive visualizations beyond text output; REPL Harness, providing a persistent Python execution environment that allows models to create and advance tools for global utilities across sessions; and the MinT training infrastructure, specifically developed for MoL, where the Actor and Training Agent exchange only adapters, not full weights, supporting millions of adapter catalogs at a trillion-parameter scale.
MindLab also created two proprietary evaluation datasets. Macaron ChatBench assesses agent honesty in long-context scenarios, checking whether the model admits knowledge gaps instead of fabricating answers. Macaron LivingBench simulates a dynamic sandbox with three interacting mechanisms: dynamic noise, a dynamic living environment, and a dynamic user, measuring the model's ability to continuously understand user needs, verify information accuracy, replan task paths, and complete tasks before user patience expires.
On standard tests, Macaron-V1 achieves a generation speed of 320 tokens per second, significantly faster than comparable models, although this may reflect low inference load during early testing. Macaron-V1 proves that Chinese artificial intelligence is defined not only by the scale of the base model training but also by the efficiency of subsequent fine-tuning. MindLab's achievement using 64 GPUs confirms the feasibility of specialization while maintaining computational efficiency. As the quality of base models converges, the competitive advantage shifts towards the post-training layer, where MoL methods may define the ecosystems attracting the most capable applications.
Huawei Cloud has introduced CloudRobo, a comprehensive platform for developing embodied artificial intelligence. This platform covers the entire workflow: from data creation and model development to robot simulation and deployment. With this move, the company has declared its position as a provider of the intelligent 'brain' for robots, rather than a direct manufacturer.
This strategy mirrors Huawei's approach in the smart car sector, which led to the success of SERES/Ask. Instead of creating the final product independently, Huawei provides the intelligent core, ecosystem, and development platform, allowing other companies to create competitive products. More than twenty companies, including Youibot and Huayan Robotics, have joined CloudRobo as initial partners.
The architecture of the CloudRobo platform is based on three main elements. First, a PB-level data foundation provides access to millions of ready-made datasets. Second, a cloud engine for model production utilizes a series of Pangu models running on Ascend AI chips. Third, an internal Real-Sim evaluation system allows for realistic virtual testing.
Developers gain access to over 20 models optimized for Ascend and can connect a robot to the cloud in mere hours, while deploying a model takes minutes. The platform covers the entire process of embodied AI development: data collection, model training, simulation evaluation, hardware integration, and operation.
The logic of this strategic decision replicates Huawei's proven model. In the smart vehicle segment, Huawei supplied intelligent driving systems, cockpits, and electric drive technology, while partners handled manufacturing. This model allowed SERES to transform from an unknown manufacturer into a premium electric vehicle brand, and the M9 model became a bestseller in the segment above 500,000 yuan. Now, Huawei is applying the same division of labor concept to embodied AI: the company provides the intelligent operating system, cloud platform, and basic AI models, while robot manufacturers focus on hardware development, production, and vertical application scenarios where they possess expertise.
The high gross profit margin of controllers at 79.8% compared to the equipment margin of 38.4% confirms the priority of the strategic focus on software-defined value.
The current time reflects industry maturity. Robot motion control capabilities are largely standardized among major manufacturers, making the intelligent processor the primary differentiator. The fragmented robotics landscape, encompassing industrial manipulators, service robots, quadrupeds, and humanoids, requires a unified development infrastructure to avoid every team recreating everything from scratch. Huawei CloudRobo solves the problem of manual development, where individual groups wasted excessive resources on creating custom data pipelines, model training, simulation environments, and hardware integration instead of innovating in applications.
Although it remains unclear whether Huawei can replicate its smart vehicle ecosystem strategy in embodied AI, given that cars feature high standardization and uniformity, whereas robots have extremely diverse requirements, Huawei already possesses the M-Robots OS based on OpenHarmony and over 15 embodied AI startups founded on talent nurtured by Huawei, indicating the company's influence extends far beyond the CloudRobo platform itself.
