2026 China AI Large Model Platform Ranking – April

Analysis by: Canshang Ximei Juice

Editor: Xingnai Media | AI Large Model Factory

Trends in Domestic Large Model Development

In April, domestic large model products for consumers entered a new phase of “tiered payment,” but market feedback showed significant differentiation. Doubao has sparked discussions on social platforms regarding the shift from free to paid services, with some users expressing that the standard version priced at 68 yuan/month is too high, while the professional version at 500 yuan/month has been humorously compared to hiring an assistant directly. This stark contrast in opinions indicates that the commercialization of large models cannot simply follow the “free to paid” logic typical of internet products. Users are willing to pay only when they clearly perceive the unique value they receive from the paid service.

Doubao’s three-tier pricing strategy implicitly aims to filter users by capability. The standard version at 68 yuan/month targets light productivity scenarios, such as PPT generation and meeting minutes organization. The enhanced version at 200 yuan/month caters to advanced users, allowing for data analysis and multi-turn deep reasoning tasks. The professional version at 500 yuan/month focuses on enterprise-level users or professional creators, offering exclusive model fine-tuning and priority inference channels. This “stepwise value anchoring” is a rational business design. However, the question remains: can the current model capabilities meet user expectations for the “professional version”? If users encounter issues like hallucinations or logical gaps despite paying 500 yuan/month, their willingness to pay will naturally decline.

In contrast to other players’ strategies, the differences are notable. Recently, Zhiyu chose to raise prices by 83%, yet, thanks to the robust performance of GLM-5.1 in coding scenarios, the call volume actually increased by 400%. This suggests that professional users prioritize “result reliability” over absolute pricing. DeepSeek, on the other hand, uses a “quick mode + expert mode” approach to tier the user experience. Free users can still access basic functionalities, while paid users benefit from deeper insights and priority responses. This “visible capability, optional payment” gradual design leads to lower user resistance.

It is clear that the success of tiered payments hinges not on whether money is charged but on the transparency of the tiering logic and the perceptibility of value delivery. More importantly, this round of commercialization is not just about “raising prices”; it signals the transition of large model products from traffic-focused to productivity-focused offerings. Data indicates that Doubao has surpassed an average daily token usage of 120 trillion, doubling in three months. The number of enterprise clients using over a trillion tokens on the Volcano Engine has increased from 100 to 140 since the end of last year. These figures demonstrate that users and enterprises are moving from superficial trials to high-frequency usage of large models. Once models can integrate into workflows to help users complete more complex, time-consuming, and professional tasks, paying will no longer feel like a sudden commercial action but rather a natural outcome of enhanced efficiency.

Tiered payments merely mark the starting point of commercialization, not its endpoint. The future of competition may not lie in pricing itself but in whether a precise match of “scenario-capability-value” can be established. It is crucial to ensure that users clearly perceive the value of their payments: spending money should enable them to complete PPTs faster, write code more reliably, and carry out content creation with less worry. Only then can large models escape the shallow traffic business and truly step into a mature commercial cycle characterized by subscription services, value-added benefits, and results-oriented payments.

Organizational Changes Driving AI Implementation

Recently, major domestic firms have shown a unified approach in organizational adjustments. AI is no longer merely an option for certain business units; it has been positioned at the strategic core of the group. Alibaba’s restructuring began with the establishment of the ATH business group, which aims to unify various AI applications, including Tongyi, Qianwen, cloud infrastructure, inference platforms, and business-oriented AI applications, into a more cohesive collaborative system. The Tongyi Lab has been upgraded to a business unit, with a technical committee led by Wu Yongming overseeing model development, AI cloud infrastructure, business technology platforms, and inference platform construction. This indicates Alibaba’s intent to move away from a fragmented approach to AI business and shift towards systematic collaboration centered around tokens.

Similarly, Baidu’s reform appears to push AI beyond just a technical department to a universal working method for all employees. It has abolished the previous T/P/E/M letter grading system, unifying it into a digital grading system. AI tool application, large model implementation capabilities, and AI business outputs have been integrated into the core performance assessments, with a clear requirement that core employees’ efficiency using AI tools must meet or exceed 20%. The logic behind this is straightforward: implementing large models cannot rely solely on a few algorithm teams but must reorganize AI across all business areas, including search, cloud storage, maps, and health.

Tencent and ByteDance exhibit another organizational acceleration approach by integrating AI into more real-world scenarios through product matrices. In April, Tencent advanced its capabilities by launching several projects, including QClaw, the mixed Yuan Hy3 Preview model, ima knowledge agent “Copilot,” and WorkBuddy, which connects to Tencent Docs. ByteDance rapidly pushed model capabilities into high-frequency scenarios like voice interaction, video generation, and 3D generation through Doubao, Volcano Engine, JiMeng, and automotive AI solutions. Rather than merely adjusting organizational structures, both companies emphasize accelerating model capabilities to the business forefront through faster product iterations.

The competition among large models is moving away from focusing solely on parameters and rankings, entering a system phase where organizational strength dictates implementation capability. As the technological gap narrows, the true differentiators will not only be the models themselves but also the collaborative efficiency among R&D, computational power, products, data, scenarios, and commercialization. Major companies are adjusting their structures, reconstructing assessments, and integrating platforms because AI implementation is no longer a task for a single department. Instead, it requires a reallocation of resources across the organization around intelligence.

As major companies leverage organizational advantages to amplify their technical strengths, the agility of startups remains a core competitive edge. The answer may lie in focusing on differentiation. While organizational strength determines implementation speed, the strength of scenarios determines commercial value. This could be the division of labor code for players of various sizes in the large model era.

Vehicle Applications as the Largest Ground for Large Model Implementation

During the Beijing Auto Show in April, Volcano Engine, iFlytek, and SenseTime simultaneously launched vehicle-mounted large model solutions. An evident signal is that vehicle scenarios are becoming one of the most promising avenues for large model commercialization. Automakers are willing to pay for edge deployment, low-latency responses, and multilingual capabilities because smart cabins have become essential features impacting user experience and market competitiveness.

When users can perform navigation, car control, entertainment, and work tasks via voice commands, large models evolve from mere “chat assistants” in car systems to integral components of the driving process, connecting people, vehicles, services, and scenarios into a mobile intelligent entity. Specifically, Volcano Engine employs an Agentic AI architecture to create an integrated loop of perception, reasoning, execution, memory, and learning across key functional domains, enhancing the rapid coupling of Byte products with the actual needs of car manufacturers. iFlytek positions “edge + overseas” as dual engines, with its exclusive vehicle-mounted large model SparkAuto EMM adapted to various computational platforms, reducing command response delays to 150ms while supporting 32 languages across 60 countries, directly addressing the global needs of Chinese automakers.

SenseTime’s Sage Box adopts a three-tier architecture of “edge model + thousand-machine system + native intelligent agent,” focusing on flexible empowerment for “one brain, multiple forms.” This approach integrates multi-modal comprehension and generative capabilities into the vehicle scenarios. These developments indicate a shift from “cloud-based showcasing” to “practical deployment in vehicles,” entering a true engineering phase of implementation.

Moreover, the vehicle scenario itself is becoming an “accelerator” for large model capability iterations. In driving conditions, models must handle ambiguous commands, noise-resistant voice interactions, and multi-tasking while meeting higher demands for real-time performance, stability, and safety. Compared to ordinary dialogue scenarios, the vehicular environment serves as a high-pressure test, compelling large model technology to mature rapidly. For instance, iFlytek’s edge model can conduct voice interactions and intelligent planning locally, reducing reliance on cloud networks. SenseTime’s SenseNova U1 series achieves unified multi-modal understanding and generation at 8B-MoT specifications, while Volcano Engine reinforces copyright and portrait safety guarantees in video generation capabilities.

These capabilities honed in vehicle scenarios might also benefit education, healthcare, industrial sectors, and more. However, the automotive sector is not the endpoint; “spatial intelligence” may well be the next frontier for large model implementation. Qunkong Technology, listed as “the first global spatial intelligence stock,” integrates spatial design, 3D generation, and interactive capabilities. Ant Group’s Lingguang also launched a world model experience on mobile, allowing users to upload images and generate 3D worlds within seconds.

As large models begin to understand and generate three-dimensional spaces, the relationship between people and vehicles can evolve from “commands-execution” to a more natural “intention-collaboration.” Smart cabins will no longer be just a combination of screens, voices, and applications but will evolve into a perceivable, understandable, and service-oriented mobile living space. This means that competition in vehicle-mounted large models is shifting from comparing parameters and isolated capabilities to a systematic stage involving “technology-scenario-ecosystem.” What automotive companies truly need is not just a chatty voice assistant but a comprehensive intelligent system capable of understanding driving scenarios, connecting service ecosystems, and continuously optimizing user experience. As edge deployment, multilingual capabilities, and spatial intelligence advance, large models will reshape not only human-vehicle interactions but also the entire value chain of smart mobility. This could be the true proposition behind the “integrated cabin-driving” competition.

Domestic Large Model Updates

In April, Alibaba released a series of AI products: the next-generation large model Qwen 3.6 series (topping the domestic rankings), voice model Fun-ASR1.5, image model Wan2.7-Image, and launched no-code development tools Meoo, world model Happy Oyster, digital human “Qianwen Xiaojiuwowo,” digital employee QoderWake, and DingTalk AI hardware A1 Pro. At the same time, Alibaba established the ATH business group and upgraded the Tongyi Lab, accelerating AI implementation from models to products and from organizations to scenarios.

On April 30, Alibaba launched digital employee QoderWake and its mobile version, covering both enterprise and personal scenario needs. QoderWake is the industry’s first secure, controllable, and continuously evolving production-grade digital employee product that can assume roles such as software engineer, operations, and analyst in real work settings. Currently, QoderWake is open for testing, allowing individuals and enterprises to apply for hiring one or more digital employees or customizing exclusive digital employees based on their business processes.

On April 30, DingTalk officially launched the DingTalk A1 Pro, now available for sale at the official Tmall DingTalk flagship store. This new product in DingTalk’s AI hardware family features a large 2980mAh battery, enabling continuous recording for 180 hours and standby for 180 days. Coupled with DingTalk’s AI transcription capabilities, A1 Pro can convert recordings into text in real-time, summarize analyses using AI large models, and support real-time multilingual translation. It includes over 200 AI meeting templates covering typical work scenarios such as customer visits, interviews, legal consultations, and cross-border meetings, helping users turn voice communications into structured online knowledge.

On April 22, Alibaba officially launched the ecological AI assistant digital human image “Qianwen Xiaojiuwowo,” recognized for her iconic smile. This female digital persona can engage in conversation and manage daily tasks such as ordering food, buying tickets, hailing rides, and planning schedules, executing cross-service operations with a single command. The Qianwen App is now live, with plans for gradual integration into Alibaba’s ecosystem applications, including Taobao and Fliggy.

On April 20, Alibaba’s Tongyi Lab announced the official launch of the voice recognition large model Fun-ASR1.5. This model, based on a unified architecture, seamlessly covers 30 languages, seven major dialects of Chinese, and over 20 regional accents, accurately transcribing ancient poetry recitation. Fun-ASR1.5 is now available on Alibaba Cloud’s Bailian platform, providing API services to customers across various industries, including education, media, finance, technology, and culture.

On April 20, Alibaba announced the official release of Qwen 3.6-Max-Preview. This model is an early preview version of the new generation flagship Qwen series, boasting enhanced world knowledge and instruction-following capabilities, with significant performance improvements in agent programming tasks. Users can interact with Qwen Studio and will soon access it via Alibaba Cloud Bailian API under the name qwen3.6-max-preview. According to third-party evaluation list Artificial Analysis, Qwen 3.6-Max-Preview outperformed models like GLM5.1 and MiniMax-M2.7, ranking as the best domestic model.

On April 16, Alibaba’s ATH business group launched the open-world model product “Happy Oyster,” focusing on real-time world creation and interaction. This product can generate dynamic 3D environments suitable for film production, game development, and more. Happy Oyster, part of the ATH AI innovation department, is currently in internal testing, with users able to join a waiting list through the official website. Built on a native multi-modal architecture, Happy Oyster supports multi-modal input and audio-video joint generation in streaming world models.

On April 15, Alibaba’s ATH business group released its first AI development tool, Meoo, integrating four leading models: Qianwen, Kimi, GLM, and MiniMax. It also includes core services like Alibaba Cloud database and storage. Users with no programming background can describe their ideas in natural language, and Meoo can automatically generate complete front-end and back-end websites or H5 pages within a minute, deploying them on Alibaba Cloud with a single click.

On April 14, Qianwen launched the “Table Agent,” allowing users to directly generate and edit Excel files within conversations. Users can request Qianwen to retrieve information and create tables, organize multi-turn dialogue content into tables, or generate tables based on images or documents. The system typically outputs downloadable Excel files within 1-2 minutes without needing to copy and paste again, supporting modifications through natural language.

On April 13, Alibaba Cloud announced adjustments to the free API quotas for standard and professional version users, supporting pay-per-use. The free quota for DataWorks standard version users is set at 100,000 calls/month, while the professional version will have a free quota of 500,000 calls/month. Any usage beyond these limits will be charged according to OpenAPI’s pay-per-use model. The adjustments will roll out regionally between April 14 and April 23.

On April 10, Alibaba’s ATH team stated that HappyHorse, developed by its innovative division, is currently in internal testing and will soon open its API. The ATH Innovation Division has launched a new exploratory plan for AI-era interactions, with HappyHorse being part of this initiative, with more products to be introduced soon. Currently, HappyHorse has registered an official Weibo account, indicating a formal introduction is forthcoming.

On April 9, Alibaba Cloud’s Bailian officially launched the “Memory Bank” feature. This system includes four modules: “Extraction-Storage-Retrieval-Injection.” After each interaction with the AI Agent, the system automatically extracts and stores key information based on configured memory rules. It can also trigger semantic retrieval to recall relevant memories and append them to the context for personalized answers, effectively enhancing the Agent’s long-term memory capability. The “Memory Bank” feature is currently available for free to all users, who can access it through the API or easily install it via products like OpenClaw.

On April 8, Alibaba’s CEO Wu Yongming announced the formation of a new Alibaba Token Hub (ATH) business group, requiring all associated businesses to commercialize around tokens. This organizational change was also accompanied by a personnel restructuring. Zhang Kaifu, head of the AI business in the Chinese e-commerce group, will no longer serve in that role, and the “Intelligent Search and Push Products” division has been split into “Platform Users and Products” and “Intelligent Algorithms.” The “Future Innovation Division” responsible for multi-modal tasks has merged into the ATH business group.

On April 8, Alibaba released a letter to all staff outlining the upgrade of the Tongyi Lab to a business unit, led by Zhou Jingren. Li Feifei has been appointed CTO of Alibaba Cloud, while Wu Zemeng will focus on the Group’s CTO responsibilities. The CEO position for Taobao Flash Sale has been taken over by Lei Yanqun. At the same time, Alibaba has established a technical committee, with Wu Yongming as the leader and members including Zhou Jingren, Wu Zemeng, and Li Feifei. Zhou Jingren serves as the chief AI architect of the technical committee, while Li Feifei oversees Alibaba Cloud technology and AI cloud infrastructure development. Wu Zemeng is responsible for the construction of business technology platforms and AI inference platforms and acts as the convenor of the technical committee.

On April 3, the globally recognized large model blind test ranking list LMArena’s Code Arena released its latest rankings, where Alibaba’s newest large language model Qwen 3.6-Plus ranked second globally, surpassing giants like OpenAI, Google, and xAI, becoming the highest-ranked Chinese large model on the list.

On April 2, Alibaba’s new generation large language model Qwen 3.6-Plus was officially launched. Compared to the previous generation, Qwen 3.6 shows significant overall performance improvements, especially in programming, agent capabilities, and tool invocation, with enhanced adaptability to mainstream agent frameworks, unlocking new potential for the model to perform complex tasks in open environments. Qwen 3.6-Plus is already available on Alibaba Cloud Bailian, with a minimum cost of 2 yuan per million tokens. Qwen 3.6 has also been introduced to Alibaba’s AI applications and platforms like Wukong and Qianwen App.

On April 2, after its launch, Qianwen AI Glasses underwent its first OTA upgrade, introducing the initial batch of “AI Task Management” capabilities. By deeply integrating with Taobao Flash Sale and Alipay, it supports high-frequency daily services such as phone bill recharging, bike scanning, parking payments, and voice-ordered food delivery, demonstrating that AI glasses can transition from “answering questions” to “getting things done,” serving as a physical interface for large model applications in the real world.

On April 1, Alibaba’s image generation and editing unified model Wan2.7-Image was officially launched. Addressing pain points such as aesthetic fatigue and color control in current AI-generated images, Wan2.7-Image delivers more “human-like” character generation, precise color control, and long-text rendering capabilities.

On April 29, Tencent Docs announced its formal integration with the WorkBuddy database. Users only need to authorize once on the WorkBuddy PC or WeChat mini program, allowing the AI to access materials in Tencent Docs without needing to download, upload, or frequently switch applications, automatically completing the entire process from material retrieval to content production.

On April 28, Tencent Cloud hosted a city summit in Chongqing, announcing the upgrade of its full-stack enterprise-level agent product capabilities, launching several new products including ClawPro dedicated cloud version, ADP intelligent workstation, Agent Memory, and Agent Storage, and signing cooperation agreements with over ten government agencies and enterprises.

On the same day, Tencent Cloud announced adjustments to the billing plans for CodeBuddy and WorkBuddy. The enterprise flagship version was renamed “SaaS Enterprise Version,” with a price adjustment from “78 yuan/person/month” to “198 yuan/person/month.” The enterprise exclusive version was renamed “Dedicated Cloud Enterprise Version,” with a price increase from “158 yuan/person/month” to “316 yuan/person/month.” The new billing plans will take effect from May 15, 2026.

On April 23, Tencent’s mixed Yuan released and open-sourced the Hy3 preview language model. This model employs a MoE architecture with a total of 295 billion parameters and supports 256k context. The official statement claims it significantly enhances performance in inference, code, and agent dimensions. The Hy3 preview has been launched across various core Tencent products, including Tencent Cloud, Yuanbao, QQ, and Tencent Docs, and supports integration with mainstream agent frameworks like OpenClaw and KiloCode.

On April 16, the mixed Yuan 3D world model 2.0 (HY-World 2.0) was officially released and open-sourced. This multi-modal world model can automatically generate, reconstruct, and simulate 3D worlds based on different input types such as text, images, and videos, while supporting the export of various 3D asset formats (Mesh/3DGS/Point Cloud) for seamless integration into existing game workflows for rapid generation of game maps and level prototypes.

On April 9, Tencent announced the launch of QClaw V2, which implements three core capabilities: multi-agent, application connectors, and Lobster Butler, allowing users to create multiple agents, each customizable with different expertise, skills, and permissions. It also connects to numerous third-party applications, claiming that single-task operation steps can be reduced by over 60%. To address security concerns, QClaw introduced the “Lobster Butler” feature, enabling users to activate a secure protection environment with one click, intercepting malicious prompts, skill poisoning, accidental file deletions, and sensitive information leaks.

On April 8, QQ Browser officially released its first browser “Lobster”—QBotClaw, allowing users to freely configure API keys for mainstream large models in China. It integrates QQ Browser’s Skill, enabling users to ask questions directly and access the “Lobster” feature. The Mac version was launched first, with a Windows version expected soon. QBotClaw features deep memory capabilities and can connect directly to WeChat.

On April 3, Tencent Cloud officially released the “Lobster” memory service, TencentDB Agent Memory. Developed by the Tencent Cloud database team, this memory engine builds a four-layer progressive memory system from raw dialogue to user profiles. Evaluation data shows that after integrating this service, OpenClaw’s overall answer accuracy reached 76.10%, nearly 59% higher than the original memory performance. Currently, Agent Memory is seamlessly integrated into Tencent Cloud Lighthouse, ClawPro, and other products in the form of a plugin, supporting easy activation for free.

ByteDance has been rapidly iterating on AI models and applications, launching the full-duplex voice large model Seeduplex, the 3D generation model Seed3D 2.0, and the high-definition video generation Seedance 2.0 API, alongside AI cabin solutions and the JiMeng “Little Octopus” tool. Doubao’s daily token usage has surpassed 120 trillion, and the service is now testing paid features for complex tasks.

On May 4, Doubao’s App Store page displayed a paid version service notice. In response, the Doubao team affirmed that the app will continue to provide free services while exploring additional value-added services, with details of the related plan still under testing. Sources close to Doubao indicate that the paid features will primarily focus on complex tasks and productivity scenarios, such as PPT generation, data analysis, and film production. As model capabilities continue to upgrade, the product is becoming capable of meeting increasingly complex high-value tasks. However, these tasks require more computational power and inference time, leading Doubao to plan to introduce paid services to meet the demands of these complex scenarios, while the free version will still cater to everyday user needs.

On April 24, the opening day of the Beijing Auto Show, Volcano Engine released its new generation automotive AI solution based on the Agentic AI architecture, including AI cabin suite solutions and the Doubao cabin assistant solution. This new solution fundamentally disrupts the previous generation of smart cabin voice assistant architecture, moving from a “intent domain + multi-agent collaboration” model to an integrated loop of perception, reasoning, execution, memory, and learning through one AI brain linking the entire vehicle, covering core functionalities such as vehicle control, intelligent driving, navigation, and cabin service.

On April 23, ByteDance officially launched the next generation 3D generation model Seed3D 2.0, achieving state-of-the-art results in geometric precision and texture material generation. This model employs a coarse-to-fine two-stage strategy and MoE architecture, with a preference rate of 98.3% in geometric generation and over 69% in texture generation based on blind evaluations by 60 3D modeling professionals. The model now supports component segmentation, articulated assets, and scene composition generation. The Seed3D 2.0 API service is now live on Volcano Ark.

On April 21, at the AI Innovation Tour in Chengdu, Volcano Engine announced that the Seedance 2.0 API now supports native 1080P full HD video generation without requiring post-processing super-resolution, significantly enhancing detail and light-shadow layers. This capability is applicable to various creative scenarios in film, advertising, and e-commerce, efficiently driving the transition from content generation to deliverable output, providing more efficient and higher quality content creation support for enterprises and creators.

On April 9, ByteDance announced the launch of the native full-duplex voice large model Seeduplex. Compared to the previous half-duplex Doubao end-to-end voice model, Seeduplex offers an enhanced interaction experience with improved naturalness and smoothness, now fully integrated in the Doubao App. On the same day, Doubao announced an upgrade to its calling capabilities, with real-time voice calling now integrated into the full-duplex voice large model Seeduplex. As a native full-duplex end-to-end voice large model, Seeduplex can achieve precise anti-interference and dynamic stop detection in complex acoustic environments, offering a smoother and more natural voice interaction experience. Following its launch, Doubao’s voice calling has further improved in dialogue naturalness, response speed, and anti-interference performance, making conversations feel more seamless.

On April 9, JiMeng AI launched its first collaborative AI narrative creation tool “Little Octopus,” introducing the innovative Vibe Create creation model. As an exploratory attempt, Octo’s functionality is not yet fully online, with open beta applications available only on the JiMeng Web platform. In terms of interaction, Octo supports a “dialogue + multi-modal mixed” co-creation approach, with its intelligent agent actively engaging with creators through images, audio, and other forms, perceiving interface content and user actions in real-time, enabling asynchronous parallel creation.

On April 2, Volcano Engine announced at the AI Innovation Tour in Wuhan that the Seedance 2.0 API is open for public testing for enterprise users. Volcano Engine has established industry-leading copyright and portrait safety guarantees for Seedance 2.0, covering various modalities involved in video generation throughout the entire creative process and detecting and defending against infringement and deep forgery, effectively protecting the rights of copyright holders and creators. As of March this year, Doubao’s average daily token usage has surpassed 120 trillion, doubling in the past three months and growing 1000 times since its release in May 2024. Currently, the number of enterprises using over a trillion tokens on the Volcano Engine has increased from 100 at the end of last year to 140.

On April 28, Baidu announced the initiation of a company-wide reform of its job grading system, officially executing a unified adjustment from the existing T/P/E/M letter grading system to a digital grading system from level 5 to 12 starting May 1, 2026. Simultaneously, AI tool application, large model implementation capabilities, and AI business outputs will be included in core performance assessments for all staff, with a clear requirement for key employees to achieve a 20% improvement in AI tool efficiency, making AI core capabilities a hard criterion for employee promotions and evaluations, fully aligning with the company’s all-encompassing AI transformation strategy.

On April 27, during Baidu AI Day, Baidu Wenku and Wangpan jointly launched the general intelligent agent GenFlow 4.0 and upgraded the Office Agent, allowing users to deploy OpenClaw in the Wangpan general intelligent agent GenFlow 4.0, transforming the Wangpan into their own “AI workspace.” To date, the monthly active user count for GenFlow 4.0 has reached 100 million, with a monthly task delivery volume of 200 million.

On April 24, Baidu officially introduced the upgraded AI search engine at the 2026 Baidu Creator Conference. The upgraded AI search engine centers around the Master Agent, adding a demand planning agent and an organizational generation agent capable of autonomously breaking down complex tasks, invoking tools, and linking services, transforming the search experience from “find and know” to “get and do.”

On April 15, the Wenxin large model team at Baidu officially open-sourced the text-to-image model ERNIE-Image. This model is based on a single-stream Diffusion Transformer (DiT) architecture and includes a lightweight prompt enhancer to expand brief inputs into richer, more structured descriptions. With a parameter scale of only 8B, it can be run on consumer-grade graphics cards with just 24GB of memory, generating results comparable to top commercial models.

On April 13, Baidu Intelligent Cloud’s DuClaw announced a new round of upgrades. Following previous integrations with Xiaodu hardware, it further integrated with Baidu Maps, enabling users to deploy Xiaolongxia with a single travel command, synchronizing Xiaodu and Baidu Maps to complete itinerary queries, weather conditions, traffic planning suggestions, and provide real-time feedback to users via Xiaodu hardware.

On April 2, Baidu Health announced the official launch of its AI product “You Medical Assistant,” designed for doctors. This is the first full-scene doctor working platform in China that deeply integrates authoritative medical searches with task-based AI execution. “You Medical Assistant” includes two core modes: search and task, with the search mode supported by tens of millions of medical data for authoritative evidence sourcing, aimed at creating a “Chinese version of OpenEvidence.” The task mode, based on the Claw framework, can autonomously complete complex tasks such as academic research, paper writing, and patient follow-ups, marking a leap from “information retrieval” to “task completion” in medical AI. Baidu Health also revealed that the search mode experience for “You Medical Assistant” is now available on the App, with gradual access to the task mode.

On April 20, “The Dark Side of the Moon” officially released and open-sourced its latest model Kimi K2.6. The model reportedly achieved uninterrupted coding for 13 hours during tests, with significant upgrades to its agent cluster architecture, supporting 300 sub-agents to collaboratively complete 4000 cooperation steps. In numerous benchmark tests, it performed comparably or better than closed-source models such as GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro, making it the strongest coding model yet from “The Dark Side of the Moon.” The model is now available on kimi.com, Kimi API, the latest Kimi App, and Kimi Code programming assistant, accessible to all users.

On April 27, Ant Group’s Lingguang App officially launched the “Experience World Model” feature, becoming the industry’s first AGI product that can be experienced on mobile. Users can upload an image and generate a 3D world on their mobile devices within seconds, exploring it freely for up to 60 seconds. This feature is powered by the Ant Lingbo LingBot-World-Fast world model, which is now open-sourced. The Lingguang App manager, Cai Wei, stated that this is another practice in exploring the boundaries of intelligence, aimed at providing a great AI experience for everyone. Users can now download the Lingguang App from major app stores to experience it.

On April 22, Ant Baijing officially launched Ling-2.6-flash, an instruct model with a total parameter count of 104B and 7.4B activated parameters. This model focuses on “Token Efficiency,” being faster, more economical, and more suitable for large-scale real-world applications while maintaining competitive intelligence levels. According to evaluations from third-party authority Artificial Analysis, Ling-2.6-flash exhibited outstanding token efficiency, achieving an Intelligence Index of 26 with 15M output tokens while controlling output consumption at relatively lower levels compared to some models reliant on more resources.

On April 20, Ant Lingguang released a new generation of flash applications called “Lingguang Circle,” aimed at creating accessible consumer-level Coding Agents. Building on the existing “30-second application creation” feature, Lingguang Flash Applications continue to enhance multi-agent collaboration, multi-modal generation, and mobile native capability integration, making it the first platform to allow users to create, distribute, use, and iterate AI applications on mobile devices using natural language, achieving personalized creation with zero coding, zero deployment, and zero barriers. To date, Lingguang users have created over 30 million flash applications.

On April 16, Ant Lingbo Technology announced the open-sourcing of its streaming 3D reconstruction model LingBot-Map. This model requires only a standard RGB camera to estimate camera pose in real-time during video capture, reconstructing the 3D structure of scenes, providing continuous, stable, and real-time spatial perception and understanding capabilities for applications like robotics, autonomous driving, and AR glasses. The model supports approximately 20 FPS real-time inference and can run continuous long videos without significant accuracy degradation, balancing precision, speed, and long-term stability.

On April 2, Ant Data Science’s specialized lobster product DTClaw entered internal testing. Unlike current market offerings that only handle document formatting, meeting minutes, and data collection, DTClaw is positioned as a “professional lobster,” offering 24/7 online AI agent services tailored for financial experts, investment advisors, data specialists, and others. DTClaw is equipped with hundreds of professional skills and pre-set templates covering high-value scenarios in finance, investment, data analysis, and research and testing, allowing users to deploy it with ease, avoiding the need for extensive training.

On April 16, MiniMax launched the world’s first cloud-based sandbox Hermes – MaxHermes. MaxHermes integrates the learning loop and self-evolution capabilities of the Hermes Agent with the MiniMax M2.7 model, allowing users to have an AI agent that becomes increasingly knowledgeable over time without local deployment, available within 10 seconds in the cloud. It is reported that MaxHermes differentiates itself through its unique learning loop mechanism, automatically refining reusable skills from complex tasks and saving them as independent documents, which can be loaded as needed for continuous improvement based on new usage feedback.

On April 12, MiniMax M2.7 was officially open-sourced globally, integrating with leading domestic and international chip manufacturers and inference platforms, completing model access and inference adaptation work on the first day of open-sourcing.

On April 10, MiniMax officially released a new generation of music generation model Music 2.6. This update has achieved comprehensive evolution from the underlying engine to creative tools, significantly improving generation latency, music control, and acoustic quality, introducing a brand-new “Cover” creation feature and Music Skill aimed at the AI Agent ecosystem, with a 14-day free trial for global creators.

On April 9, MiniMax announced the release of MMX-CLI, a command line tool for AI Agents. With MMX-CLI, agents can natively call MiniMax’s latest models for programming, video generation, voice synthesis, and music creation in environments like ClaudeCode and OpenClaw, eliminating the need to adapt cumbersome interfaces or write additional MCP servers.

On April 24, iFlytek officially showcased a new generation multi-modal intelligent cabin solution, developed from the iteration of the Spark large model, at the 2026 Beijing International Auto Show. This cabin system achieves ultra-fast interactive responses, with command intent response delays as low as 150ms, and recognition accuracy exceeding 90% for conversationally ambiguous commands and fragmented car usage needs. The dedicated edge-mounted large model SparkAutoEMM has completed compatibility with various in-vehicle computational platforms, enabling local voice interaction, scenario services, and intelligent planning, reducing cloud dependency and enhancing driving stability and privacy security. iFlytek also displayed the overseas version of the Spark large model’s application results, fully supporting 32 mainstream languages, with service coverage extending to 60 countries and regions worldwide, having adapted and deployed several mainstream overseas automotive models.

On April 16, the 2026 iFlytek AstronClaw upgrade release conference was held, unveiling nine innovative products. Among them, the enterprise-level procurement assistant “Procurement Claw” made its debut, deeply integrating the newly upgraded i