Global AI Daily News

DeepFlow Intelligence open-sourced DojoAgents, gaining 1000+ GitHub Stars within one week. The framework decomposes market scanning, news retrieval, company analysis, industry chain mapping, portfolio construction and risk analysis into callable tools, using Agent Loop to enable continuous autonomous research and decision support, helping investors quickly organize information, validate judgments and generate traceable simulated portfolios in complex markets.

Acrab launches Agent Box personal AI center at 2026 WAIC, featuring the GΞLIX 1 platform capable of running hundred-billion parameter models locally, integrating 8MB L1 Cache with 650 TOPS AI computing power. The device enables Agent context preservation, tool invocation and multi-Agent collaboration, eliminating cloud-based Token costs. CEO Ken Phua is former Arm China co-CEO, with the company raising over $350 million in cumulative funding, positioning itself as the computational foundation fo

Hu Han, head of Tencent's Hunyuan multimodal understanding, has resigned to start a business, with his former team potentially pivoting to world models. Tian Yonglong will take over VLM research. Following Yao Shunyu's leadership, resources are being concentrated on large language models, with the Hy3 model launched in July. Tencent's 2025 capital expenditure reached 79.2 billion yuan, facing competition from Alibaba (126 billion yuan) and ByteDance (70 billion USD).

After Kimi paused new subscriptions, users can access K3 via API. Testing shows: direct API connection delivers fastest but without process feedback; integrating with Claude Code adds file operations and tool calling, yet risks failed file writes and style drift; official client performs more reliably. Key finding: the same model behaves differently across various harnesses. API migration transfers inference capability but not system prompts, tool orchestration, or delivery workflows, placing st

LPU startup YuanChuan Micro completes Pre-A-round financing of hundreds of millions of RMB, led by IDG Capital. Founded by Huawei veterans, the company develops native LPU architecture for Agentic AI era inference, featuring extreme low latency through spatio-temporal compiler and hard-pipeline architecture. Core technologies support multi-modal and MoE models with flexible deployment across cloud, edge, and endpoints. The company secured multiple financing rounds in 2026 since its establishment

Shexun Technology secures over 100 million yuan in Series B funding, bringing cumulative financing to over 300 million yuan. Its Zexing AI platform integrates geometry AI, physics AI, and manufacturing AI to assist engineers with design tasks. Contract revenue surged 70% YoY in H1, with full-year projections approaching 200 million yuan, nearly doubling last year. The company designates 2026 as its overseas expansion year, starting with Germany, aiming to land around 10 clients in H2.

GPT-5.6 Pro has disproven the 30-year-old Dinitz-Garg-Goemans conjecture. Researcher Rybin guided the AI using just 4 prompts totaling 58 English words, constructing a counterexample graph with only 7 nodes and 9 edges through exhaustive verification: the minimum cost for unsplittable flow is 60, while the conjecture assumes 58, creating an irreconcilable conflict.

Lingchu Intelligence and the PKU-Lingchu Joint Lab released and open-sourced EgoSteer, a dual dexterous-hand general-purpose manipulation large model with a full-stack system, including the EgoSmith data pipeline (processing 9.6K hours of egocentric video), world-model-enhanced VLA, and Robot-Stack. The model achieves 75% success rate across 40 tasks and can generalize to new tasks with few demonstrations. The company, founded 1.5 years ago, has completed a 2 billion RMB financing round.

Researchers, using GPT-5.6 Sol on the OpenFPM platform, successfully ran CUDA programs directly on Apple M3 Pro's Metal GPU through intermediate layers like Clang/HIP, SPIR-V, Vulkan, and MoltenVK. The same program achieved approximately 10x speedup compared to CPU sequential computation, validating cross-platform GPU computing feasibility.

Youibot's FabriVLA topped the embodied intelligence leaderboard with 90% success rate on Meta-World MT50, surpassing international models including π0. Through gating self-attention and deep-shallow visual fusion innovations, this 1B-parameter model achieves SOTA performance, proving lightweight models can attain top-tier embodied intelligence and enabling edge deployment. Paired with FabriX industrial model and Jifeng humanoid robot, which has received 4,000 orders. Youibot has accumulated 800+

Messi founded venture capital firm Play Time in San Francisco in 2022 with an initial fund of approximately $200 million. His investment portfolio includes Fei-Fei Li's World Labs (valued at $5 billion), AI data platform SuperAnnotate, robotics developer Field AI, and others, covering AI infrastructure and physical robotics sectors, breaking the stereotype of athletes only investing in real estate and restaurants.

Odin, a 00s-born founder of Valhalla, dropped out of high school and self-studied into Zhejiang University. He later worked at Nobel laureate David Baker's lab before leaving to start his own venture. Having raised tens of millions USD from Source Five Capital, Qiji, and ZhenFund with a 30-person team, his goal is to build a full-modal molecular world model to accelerate human scientific discovery a hundredfold and make treatments accessible for rare disease patients.

Anthropic launches six updates to CMA platform: skill limit increased from 20 to 500, supporting one session to run entire business units; thinking effort now adjustable across five levels for complex reasoning or token-saving on simple tasks; seed sessions create in one step with up to 50 initial events; sub-agent real-time streaming drops to thread level; 7 new webhooks cover full lifecycle of environment and memory storage; version field becomes optional. Anthropic is building agents into OS-

iFLYTEK establishes YaoFang Intelligence (registered capital 300M yuan), led by Pan Jia, releasing the iFLYTEK-Embodied-Omni technical report. The company identifies that mainstream robot brain approaches lack "body cognition," proposing a dual-core architecture (VLM+VGM+AGM) with inverse kinematics training. Benchmark shows 89.6% zero-shot generalization and 93%+ dual-arm coordination, aimed at solving the "stuck in demo" problem.

Yingmo Technology's paper cuNRTO was nominated for RSS 2026 Outstanding Paper Award. The work accelerates nonlinear robust trajectory optimization on GPU, achieving up to 139.6x speedup while maintaining 100% safety constraint satisfaction, solving the critical sim-to-real deployment challenge.

Research shows that after using AI, the proportion of people admitting "I don't know" dropped from 44% to 3%, accuracy fell from 27% to 9%, yet confidence surged from 30% to 76%. Even with financial penalties for wrong answers, people remain unable to break reliance on AI's erroneous suggestions. AI's fluent hallucinated responses erode human metacognitive ability to distinguish genuine knowledge from plausible guesses, with children facing even greater risks.

DeepSeek completes its first external financing exceeding 50 billion RMB with a pre-money valuation of approximately 367.5 billion RMB. Founder Liang Wenfeng conducted an extensive 4-hour investor meeting, presenting his systematic views on the company's vision, AGI roadmap, and open-source strategy. He stated that open-source will become a competitive advantage, domestic computing ecosystems can be established within one year, and Coding Agent is the most important direction at this stage.

Show 2 related articles

Liang Wenfeng outlines DeepSeek's strategy: avoiding 3D, video, and world models, not pursuing super apps, framing restraint as a strategic trade-off for AGI success. Open source is their sweet spot, with team stability as the core non-negotiable principle. Goal is to narrow the China-US AI gap to 6 months using a fraction of the computing resources.

DeepSeek founder Liang Wenfeng recently shared his AI development philosophy: rejecting three-year commercial decision-making horizons, adhering to a thirty-year perspective. He views open-source as a survival strategy rather than a moral choice, considers C-end traffic as secondary to larger goals, and runs his organization without KPIs, driven only by vision. He follows no industry trajectory, focusing solely on the AGI evolution主线.

NUDT researchers introduce AirZoo, a unified large-scale dataset and benchmark for aerial geometric 3D vision, spanning 22 countries with over 1.2M RGB-D frames and 6-DoF pose annotations. Built on the AirSim-Cesium-Unreal pipeline, it supports three tasks: aerial image retrieval, cross-view matching, and multi-view 3D reconstruction. Experiments show significant performance gains in localization, matching and reconstruction after fine-tuning, accepted at ECCV 2026.

Meituan releases LongCat-2.0, the industry's first trillion-parameter model trained on a 50,000-chip domestic computing cluster. Simultaneously, it launches the LongCat Top Talent Campus Recruitment plan, offering industry-leading compensation and top-tier per capita computing resources to compete for elite AI talent.

Researchers from BUPT, Tsinghua University, and Shanghai Jiao Tong University released MemSlides, topping Hugging Face Daily Papers with 400+ GitHub stars. The system uses user profile memory, working memory, and tool memory to help AI remember user revision habits, generating personalized slides from round-0 and enabling scoped modifications without disrupting the entire deck, solving the core pain point of AI PPT "changing one page breaks the whole presentation".

Kunlun Wanwei releases Matrix-Game 3.5 world model at WAIC 2026, capable of continuously generating interactive worlds based on user actions and event commands while maintaining scene structure and character identity consistency in long-term generation. Through sampling step compression, DiT inference optimization, and VAE acceleration, the model achieves 720P resolution at approximately 20FPS real-time streaming video generation on a single GPU. Version 3.5 introduces the industry's first geome

Creator Irene developed the "Style Pack Skill" tool that automatically collects and organizes movie screenshots, generating color palettes, style rules, and prompt templates. The tool categorizes screenshots by scene, with each palette fixed at 10 color slots. Available on Codex and Work Buddy, it solves the problem of inconsistent style in AI-generated content by converting vague descriptions into reusable visual specifications.

Instagram head Adam Mosseri reveals the company is streamlining its traditional 13-person product team into Pods of 6-7 members, consisting of 4-6 generalist engineers and 1 product staff. As AI tools handle code execution, engineers are shifting focus from coding to task planning and code review. Expert roles have transitioned from fixed positions to on-demand involvement. He emphasizes that human value in the AI era lies in taste and strategic judgment—those who can define clear objectives and

Beijing released "Several Measures to Accelerate Intelligent Agent-Led Development", containing 10 policies covering foundational model capabilities, harness engineering, native applications, smart terminal integration, OPC entrepreneurship, token economy, security governance, computing infrastructure, and open-source initiatives, aiming to build a new intelligent economic paradigm for the AI era and promote widespread adoption of intelligent agent technology.

HiDream.ai has completed a 1.5 billion yuan Series C funding round, bringing its total financing to over 2.1 billion yuan in the past three months, with post-money valuation exceeding 1 billion USD, achieving unicorn status. Founded by Canadian Academy of Engineering fellow Tao Mei, the company focuses on native full-modality world model technology, with its HiDream-O1-Image topping global text-to-image rankings. The investor lineup includes 20+ institutions including social security fund, state

AI company Video Rebirth releases world model Olympus, adopting a video-native approach using video as the foundational carrier for world understanding and generation. The model features real-time interactivity, physical consistency, audio-visual synchronization, and long-term memory, built on the proprietary BACH video generation engine, capable of supporting gaming agents, robotics, and autonomous driving applications, and has been deployed on AMD MI350.

Former Doubao PC head Qi Junyuan launches AI startup 'Today AI', positioning it as an OS with autonomous thinking capabilities that understands interfaces, operates devices, and executes tasks directly rather than passively answering. The project started at least three months ago, raising hundreds of millions in funding from IDG and other institutions, currently offering internal testing for Mac, iOS, and Android.

Musk announces Grok Imagine will produce a full-length AI-generated Odyssey by year-end, directly challenging Nolan's live-action version ($264M worldwide opening). A 3-minute AI short showcasing impressive results became the catalyst, demonstrating individual creators' ability to generate cinematic footage with AI. Critics note this signals filmmaking shifting from capital-intensive to imagination-intensive.

Taobao's parent company Tmall released if Studio, an intelligent agent workbench that transforms design experts into a streamlined production tool. Users simply describe their needs, and the system automatically breaks down tasks and invokes models to complete designs, supporting posters, videos, landing pages and more. Testing shows direct image generation, 4 rounds of revisions under 8 minutes, and product video creation under 20 minutes, saving 1-3 days and thousands of yuan compared to outso

Seven AI models tackled all six problems from IMO 2026, with Claude Fable 5, GPT-5.6 Sol, Kimi K3, and AxiomProver achieving perfect 42-point scores. Over seven years, only 30 out of 4,347 human contestants achieved perfection (0.69%). These AIs reached perfect scores without specialized math training, signaling AI's mastery of complex logical reasoning and its potential to assist ordinary people with tasks requiring rigorous deduction like contract review and tax planning.

Baidu unveiled Miaoda 3.5 at WAIC 2026, serving over 35 million users with nearly 200,000 daily active users. The new version introduces six features including SEO Agent, iOS packaging, and shared backend. Testing revealed: shared backend functions smoothly, SEO Agent is convenient but limited, while iOS packaging crashed with white screen errors even for PRO members. The review concludes the product direction is correct, but platform stability determines user experience, questioning the actual

Gaode releases full-stack embodied AI system ABot with five new models (N1, M0.5, ER, AgentOS, C0), achieving SOTA on 17 benchmarks. The three-layer architecture enables closed-loop self-evolution via data feedback. N1 achieves 92.9% open-world navigation success rate, M0.5 outperforms previous SOTA by 20.4% in complex tasks, leveraging 20 years of spatiotemporal data to build a complete closed loop that enables robots to evolve from capable to competent.

Kinetix AI unveiled KAIBot at WAIC 2026, featuring the world's highest 115 degrees of freedom, with a 173cm height, 70kg weight, 18,000 tactile points achieving 0.1N sensing precision, equipped with a 37-DOF dexterous hand and KAI World Model algorithm, capable of autonomously playing table tennis, real-time dialogue, and opening beverage cans.

PhyAgentOS, jointly developed by Sun Yat-sen University, Pengcheng National Lab, and X-Era Lab, has been released with full open-source access. This open-source operating foundation targets physical intelligence models and heterogeneous robots. Centered on Session management, it enables AI transition from "thinking" to "reliable execution". Tests show success rates improved to 97-99% on LIBERO, 45-89% on CALVIN long-horizon tasks, and survival days in Don't Starve increased from 1.02 to 2.10 day

At WAIC, Mech-Mind showcased its "One Brain, Multiple Forms" embodied intelligence solution, enabling diverse robots to share a unified "eyes-brain-hand" system. 3D vision perceives the environment, Mech-GPT large model enables understanding and reasoning, and dexterous hands execute actions, forming a closed-loop of perception-decision-planning-execution-feedback. The products have been deployed globally in over 27,000 units, serving 100+ Fortune 500 clients across automotive manufacturing, log

Sekai is an AI platform enabling users to generate playable mini-apps through text descriptions. The platform has accumulated over 15 million mini-apps with 200,000+ daily additions, having raised $26 million in total funding. Sekai significantly lowers the barrier to app creation, allowing anyone to rapidly produce and share interactive content, following a TikTok-style content production model.

Zhao Hao, Assistant Professor at Tsinghua University's AIR, proposes the "native 4D world model" concept, emphasizing that true world models for autonomous driving and robotics require integrating geometry, time, optics, and mechanical properties. He notes that 4D representation is more interpretable and correctable than video generation. His team has released series of成果 including Dexora (ICRA 2026 Best Paper), Trellis.2 (CVPR 2026 Best Student Paper), and OmniNWM.

The SSNA team proposes NAF, a noise adaptation framework that constructs discriminative category structures using random Gaussian noise and transfers them to real data. With only 4 labeled samples per class and ResNet-18 backbone, NAF achieves a 12.35 percentage point improvement over the ERM baseline on CIFAR-10. The paper was published at ICML 2026 with open-source code available.

Kingsoft Office launches Lingxi Pro, an independent AI office agent positioned as "your personal office assistant". In testing, it organized 100,000-word transcripts into a knowledge base, generated 8-page PPTs in about 10 minutes with multi-round editing, and automatically updated Excel metrics after data modifications. Its core feature is "knowing you": understanding user work history via memory functions, distinguishing confirmed requirements from temporary attempts. Kingsoft's strategic goal

WAIC 2026 spotlights four world model benchmarks: VBench measures "realism," WorldModelBench measures "understanding," RoboTwin 2.0 measures "execution capability," and Physics-IQ measures "prediction accuracy." Magi-1 tops Physics-IQ at 56%, while Sora 2 scores only 42%. The establishment of evaluation systems is shifting the industry from vague narratives to clearer boundaries.

At WAIC 2026, iFlytek AI University upgraded its positioning from "Learning AI" to "AI Learning", launching two features: AI Learning for intelligent learning path planning and AI Vault for personal knowledge management. Experts noted that in the AI era, knowledge is becoming less valuable, while experience, judgment, and execution matter more. They recommend building a knowledge framework, becoming AI native, and taking swift action.

Supermicro released FusionOne AI at WAIC 2026, launching the desktop-class TokenBox™ and dedicated Token Factory solution to convert computing power into effective Tokens. China's daily Token usage surged 1000x to 14 trillion in two years, with AI Coding consuming over 50% of global Tokens. Supermicro's self-developed AI Coding delivered 500K lines of code in three months, validating the Token Factory concept.

It Shi Zhihang released AWE 3.5, the embodied native base model at WAIC 2026. Trained on over one million hours of human-centric data, this is the first native model in embodied AI achieving full pre-training to post-training paradigm integration. With its One Model architecture, it can simultaneously output actions and predict future outcomes, marking comprehensive release of embodied Scaling capabilities.

Zhejiang University and NVIDIA introduce Light Interaction, a training-free method that accelerates video world model inference by up to 2.59x. By dynamically managing context based on interaction states, reusing denoising cache during revisits, and leveraging software-hardware co-design with 3D sparse attention, the approach reduces latency from 228s to 88s on HY-WorldPlay while cutting peak memory from 76GB to 55GB.

Zhipu AI's Chief Scientist Tang Jie has teased a "Major Plus" upgrade for GLM. Their paper IndexCache, co-developed with Tsinghua University, was accepted by COLM. The technique optimizes long-context inference via cross-layer index reuse, cutting up to 75% of indexer computation and boosting first-token speed by 1.82x at 200K context. Zhipu also launched a 1GW domestic AI computing center and acquired Zhongke Jiahe to complete its full stack from chips to models.

Xingyuanzhi introduces DA-Nav, an orientation-aware visual language navigation framework that unifies navigation commands, spatial understanding, and yaw recovery into a single decision chain. The system achieves 98.15% deviation correction success rate, enabling robots to actively correct deviations and return to correct routes. Compared to 15.46% success rate after removing recovery data, this validates the necessity of incorporating error states into training, with the system completing over

A research team from Nanjing University, Shanghai AI Lab, NTU and Peking University releases VideoChat3, a 4B-parameter multimodal large model for general video understanding. It features I3D-ViT visual encoder achieving 16x spatiotemporal compression and adaptive frame resolution for streaming video, significantly reducing visual tokens and computational costs in long-video reasoning. VideoChat3 achieves 70.1 on Video-MME, outperforming Qwen3-VL-4B on 18 of 19 benchmarks. Model weights, code an

Sudu Technology made its domestic debut at WAIC 2026, showcasing robots based on the Sim-to-Real approach with software-hardware co-design and hierarchical architecture: upper layer for task planning, lower layer for object manipulation. The R1 robot achieved 98% zero-shot general grasping success rate. Co-founder Su Hao, a core author of ImageNet, applied 2D dataset methodology to 3D to address data scarcity in robotics. CEO Han Zheng predicts Silicon Valley will return to hierarchical architec

OpenAI plans to invest over $30 billion to build a data center in Georgia, securing 3.2GW of power supply, with operations expected to begin in 2028. OpenAI has committed $20 billion investment in the county and is seeking partners for joint financing. As part of the Stargate initiative, this move aims to support long-term AI development goals.

Jensen Huang backs Chinese open-source AI models in Axios interview, saying American companies should be allowed to use excellent open-source models and free AI benefits chips and data centers. Musk retweeted in support. The White House accuses Kimi K3 of distilling Claude, while 209 startups jointly oppose banning Chinese open-source models.