World Models Top Benchmarks, Deployment Platform Goes Open Source: AI Competition Shifts to Physical World and Production Environments

Recently, the AI industry's focus is no longer just model parameters and leaderboard rankings, but whether models can understand the physical world, whether they can be deployed at low cost, and how data and content boundaries are defined once they enter real business. UBTECH and Wuyuan Technology are entering through world models, while Paradigm Intelligence and NetEase Youdao are pulling competition toward deployment platforms and voice Agents. Discussions around DeepSeek, Meta, Nvidia and AI-driven sales sketch out the efficiency and governance challenges of the adoption phase.

World models race for physical common sense

UBTECH Thinker-WM took first place in all three embodied world model benchmark evaluations. One composite score was 0.740, with 3D consistency at 0.531 and interaction fidelity at 0.765, and it ranked first among all evaluated models on two individual metrics; EWMBench overall score was 4.66, surpassing mainstream models such as Qwen-RobotWorld and Sora2; on the simulation environment benchmark WorldArena Track-1, the model led across 12 key evaluation dimensions including visual quality, content consistency, physical plausibility, 3D geometric accuracy and instruction following.^[1]

Unlike previous 'stitched' approaches combining video generation, policy networks and geometric prediction, Thinker-WM is built around a unified dynamic Transformer (UDT), encoding RGB images, metric depth and robot actions into the same shared representation space, jointly evolving them through shared parameters, and then separately reconstructing video, depth maps and robot actions. Depth is no longer merely an auxiliary supervision signal, but directly shapes the shared representation as a native generative modality; in a RoboTwin single-task ablation experiment, introducing the depth modality raised manipulation success rate by 15 percentage points over a no-depth baseline. The same set of weights supports six capability modes, including joint video-action generation, world prediction, policy reasoning, video-to-action and inverse dynamics.^[1]

Another world model startup thread comes from Haixu Wu, a winner of Tsinghua University's Special Scholarship. QbitAI learned that he started a world model startup this year; his company Wuyuan Technology has completed multiple funding rounds, with a valuation reaching hundreds of millions of dollars, and investors including Ant Group, Meituan, Fortune Capital and Xinglian. His GeoPT work proposed a synthetic dynamics pretraining paradigm that 'lifts' static geometry into dynamic space, allowing models to learn particle trajectory evolution on unlabeled data and thereby acquire physical intuition. A person close to Wuyuan Technology told QbitAI that the approach continues the GeoPT research line to some extent. Haixu Wu has proposed long-sequence forecasting work such as Autoformer, and since 2025 has been doing postdoctoral research at MIT CSAIL; his advisors include Mingsheng Long and Wojciech Matusik, and the paper authors also include Kaiming He.^[6]

Deployment platforms and voice entry points compete for 'usability'

On October 16, Paradigm Intelligence and Ruisizhilian announced the launch of one-stop production-grade AI deployment platform 'ModelSphere', free and open source for enterprises under the Apache 2.0 license and available on GitHub. The platform integrates compute management, model deployment and runtime optimization into a unified process, with built-in automatic elastic scaling, PD-disaggregated deployment, three-tier KV cache, high-availability load balancing and automatic continuation after service interruption. Its AutoTune intelligent model scheduling strategy and hardware-acceleration-aware adaptive load balancing scheduling raised the system's peak throughput efficiency by about 285% over baseline, with performance jumping to 4 times the original. The platform also integrates the EngineX inference engine, supporting unified and hybrid scheduling of domestic compute including Huawei Ascend, Iluvatar CoreX, Moore Threads and Hygon DCU, and has adapted mainstream models such as DeepSeek-V4/V4-Flash, Qwen, Kimi, MiniMax and GLM.^[2]

NetEase Youdao, meanwhile, is pushing voice Agents into user co-creation. NetEase's first AI-native voice Agent, 'Bage Shuo', launched a 'user honorary co-creation plan', investing 100,000 yuan to solicit product suggestions, with 11 places and a maximum of 50,000 yuan for a single person, with no education or major requirements. Users can submit suggestions on product features, usage scenarios, experience problems and future capabilities; selected participants receive a co-creation appointment letter, a customized NetEase commemorative work badge and a direct interview pass, among other benefits, and contributions are traceable, attributable and permanently retained. At the product level, Bage Shuo uses self-developed ASR and semantic understanding pipelines to convert natural speech into structured, directly reusable high-quality text; the efficiency ratio of spoken output versus keyboard typing rises to 5 times, it supports real-time voice translation across 124 languages, translation accuracy in professional fields such as economics, physics, chemistry and medicine can reach 98%, and all data is processed locally and is free for life.^[5]

Cost, terms and governance in the deep end of adoption

On October 17, DeepSeek V4.1 Flash was updated to V4.1-Flash-1017, with simultaneous changes to the official web, APP and API models. The company said the new model achieved substantial improvements across benchmarks, while task runtime and output length were shortened, allowing users to complete tasks at lower cost and faster speed, reaching a new performance-cost Pareto optimum under multiple Harness setups; the optimization also focused on platform frameworks such as DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. Model iteration thus places unit task cost and response speed on par with benchmark scores.^[4]

On the same day, new Meta Business Suite terms sparked controversy. The new rules will officially take effect on October 30: when a business opens a Facebook or Instagram business account, it automatically grants Meta the right to scrape public content from its own website, with no separate opt-out channel; even if the business website blocks web crawlers via robots.txt, it cannot prevent content collection authorized under the business account agreement. Meta explained that scraping web content is intended to improve the accuracy of Meta AI answers, and that since businesses use the platform's business services, they should open up the right to use web content. Publishers and small and medium-sized enterprise associations criticized the rules as forcing businesses into a binary choice: either give up Meta business marketing tools or hand over website content for AI training, bypassing normal copyright negotiation processes.^[4]

Nvidia CEO Jensen Huang said in a Bloomberg interview in London on October 14 that current AI demand far exceeds supply, and he does not believe the industry is in an 'AI bubble' stage at present, though he acknowledged that supply could exceed demand in the future. He said: 'Bubbles happen when supply exceeds demand. Eventually supply and demand will balance, and at some point supply may exceed demand, but not today.' Facing market concerns triggered by AI cloud service provider CoreWeave, Huang said there is no need to rescue the company and its financial condition will be fine.^[4]

Alongside the industry's efficiency narrative, AI abuse has also produced more concrete business. Short video platforms have seen a large number of accounts named 'Xiaohai', 'Xiaotao' and 'Xiaoyu' attracting middle-aged and elderly users with similar stories such as a father falling ill and falling into debt, and selling goods to repay debt, calling viewers 'auntie' and 'mom' while pitching products. According to reports, behind this are batch-replicated AI sales actors; one account operator said the model and scripts are public and available to everyone, and that choosing the middle-aged and elderly track can earn two or three thousand yuan a day at the low end. After a batch of 'AI Wang Jianguo' accounts were shut down by platforms and related videos cleared, operators pushed supporting characters such as 'son Xiaohai' to the front, trying to create another traffic entry point.^[4]

Common trend: from 'can demo' to 'can deliver'

Putting these threads together, the yardstick for AI competition is changing. World models compete on unified modeling of 3D geometry, action consequences and physical consistency; deployment platforms compete on heterogeneous compute scheduling, inference cost and stable operation; voice Agents compete on whether they can convert high-frequency speech into reusable text and get close to real needs through co-creation mechanisms. At the same time, DeepSeek's iteration emphasizes performance-cost Pareto optimality, while Meta's terms controversy, Huang's judgment on supply and demand, and the chaos of AI sob-story selling show that the deeper adoption goes, the harder it is to bypass cost, data authorization and content governance. Worth watching next: whether world models can verify generalization on more real robot tasks, whether open-source deployment platforms can form a stable ecosystem in domestic chip clusters, and whether platform terms and governance of AI-generated content will move from corporate self-regulation to clearer rules.^{[1][6][2][5][4]}