Jina AI releases jina-ocr-v1, a compact 3.4B parameter document parsing model with only 570M activated. It achieves 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, setting new SOTA for its class. Throughput reaches 2.57 pages/sec on single A100, ranking first among 14 tested systems; FastMTP speculative decoding delivers nearly 2x speedup on L4 GPU.
Where Agents & Humans discover newsLive
Global AI Daily News
Covering 1125 days, 共 29244+ articles built agent-first, open to all tech stacks.
Send this prompt to your Agent — it will automatically read the docs and configure the skill:
Read{domain}/ainews/SKILL.md ...35
Alibaba releases Qwen3.8-Omni-Flash, a native full-modal model supporting text, image, audio, video input and 1M context window. It achieves over 26% improvement across 30 benchmarks with significantly enhanced audio-video agent capabilities and API price reduction exceeding 98%. The model can plan tasks, invoke tools and complete creations, advancing full-modal intelligence into professional production workflows.
ArcLoop launches a full-stack AI storytelling engine enabling creators to manage characters, scenes, visual styles, and assets within a unified project. Its IP World-centric architecture allows character profiles and worldbuilding elements to be reused across episodes, eliminating redundant descriptions and preventing narrative drift. The startup has closed its seed round, with founder Qin Lu previously serving as Data Science Lead at ByteDance. Similar offerings like Plot Party are also emergin
Anthropic launches the Life Sciences Verification Program beta, marking the first unlock of the Mythos model for certified scientists. The program offers two license types: standard authorization for daily research work with annual renewal, and high-risk authorization removing all safety restrictions, project-based with six-month renewal. The system adopts an offline monitoring and shared responsibility model with 30-day data retention, explicitly excluding data from model training.
Google's Gemini 4 Pro appears to have secretly entered LLM Arena under the guise of "gemini-3.8-flash", with benchmark tests comprehensively outperforming GPT-6 Astra and Claude Fable 5.1, showing outstanding results in coding, agent, and reasoning domains. Performance: DeepSWE 88%, GDPval 2064 Elo, Terminal-bench 95.3%, OSWorld 86.8%. Pricing at $2.25/M input tokens and $11.25/M output tokens. Reportedly, Google has achieved RSI (Recursive Self-Improvement) technology.
Sihao Huang, with dual MIT degrees in physics and electrical engineering and prior White House AI policy advisory experience, joins Anthropic as Head of Frontier Compute Strategy, overseeing compute infrastructure expansion, industry partnerships, and long-term planning, partnering with cofounder Tom Brown. He is the highest-ranked Chinese executive at Anthropic.
Samsung, in collaboration with Oxford University and Peking University, introduces TrOPD (Trust Region On-Policy Distillation), which enables on-device small models to stably inherit large model reasoning capabilities by identifying reliable supervision regions from the teacher model. On Qwen3-SFT-1.7B, TrOPD achieves +3.34, +4.00, +5.11, and +6.18 point improvements across math, code, instruction following, and STEM benchmarks respectively, outperforming EOPD, REOPOLD and other existing methods
ByteDance releases Doubao 2.1 Pro (0915), featuring multimodal coding capabilities that can generate code directly from design mockups, diagrams, and videos. Now available on Volcano Ark and Doubao Workbench, with over 30% reduction in token consumption for image and video reasoning. Supports converting architectural plans to 3D models and generating interactive web pages from hand-drawn sketches.
Researchers from Peking University and Yixin AI Lab, published at EMNLP 2026, discovered the "Readout Bottleneck" phenomenon where LLMs encode correct answers in hidden states but accuracy drops from 83% to 33% after vocabulary projection. A simple two-parameter unlabeled prior correction boosts Qwen3.5 accuracy from 33% (random guessing) to 57%-67%, demonstrating that model reasoning capability is separate from output expression.
BMBAI and collaborators introduce SimpleMemVLA, using visual history as direct context for robots to remember past events, achieving SOTA on four memory benchmarks. With 126-second history window, task success rate reaches 63.6%, outperforming the previous best model by 17.4 percentage points. Streaming inference reuses historical computations, reducing decision latency from 1.02s to 0.68s, with 245k token context streaming at just 1.18s. The method also exhibits visual in-context learning, allo
Mecka AI, a startup specializing in collecting human motion data to train humanoid robots, is raising a new round of financing led by Sequoia Capital at a valuation of $500 million. The company pays people to wear body sensors and use smartphones to record daily tasks like making coffee or repairing cars. Founded in 2024 by a four-person team without robotics background, Mecka projects reaching $100 million in annualized revenue by end of 2026.
Zhipu releases latest RSI progress: GLM-5.3-powered Infra Agent completed model adaptation to production-ready deployment in under two weeks on a 100,000+ domestic chip cluster, achieving 3x end-to-end throughput improvement. Infrastructure optimization traditionally requiring weeks by senior teams is now accomplished with AI participation.
ByteDance upgrades its AI education product "Doubao Love Learning" to version 2.0, powered by Doubao LLM, covering K12 subjects including Chinese, Math, and English. The app integrates AI Q&A, 1V1 interactive tutoring, photo-based problem solving, and custom AI courses, providing learning support like whiteboards, quizzes, and error notebooks to enable full-process learning. Its overseas version Gauth has exceeded 200 million downloads.
Anthropic reveals Claude generates 80% of company code, with engineers delivering 8x more per quarter. However, AI's high-frequency small PR submissions and 10x test case growth caused CI tasks to surge 25x within six months, nearly crashing the system. Three patches all failed before a distributed stateless architecture resolved the bottleneck.
WPS Comate by Kingsoft helps enterprises organize documents, data and experience into AI-readable context, enabling AI to truly enter manufacturing production environments. At CSSC Power, it manages nearly 500K documents; at HPWCH, it digitizes 200+ design specifications; at Chery, it supports 60K employees with 4,000+ digital employee applications, saving over 30M yuan annually and cutting overseas aftersales query time from 10 minutes to 1 minute. Enterprise AI competition is shifting from mod
As embodied AI robots roll out in volume, the industry's focus shifts from manufacturing robots to capability mass production. Baidu Intelligent Cloud enables rapid skill validation and replication through full-stack AI infrastructure, supporting over 50 embodied AI companies. Its Dongguan training facility covers 13 industrial scenarios with 50 heterogeneous bodies, achieving over 99.5% effective training duration.
Alibaba DAMO Academy's general medical imaging model DAMO RADAR was published in Science and open-sourced. Focusing on abdominal contrast-enhanced CT, it achieved an average AUC of 0.913 across nearly 40,000 systematic evaluations covering 18 anatomical structures and 146 clinical imaging findings, reaching radiologist expert level for the first time.
Kimi releases Kimi Code desktop app, natively compatible with macOS and Windows, featuring flagship model Kimi K3 with third-party model support. The installer is 143MB with three subscription tiers (99/199/699 CNY/month). Practical testing shows smooth Computer Use performance, built-in browser, data plugins, and screen annotation features.
Kimi releases financial industry solution integrating 10+ authoritative data sources, 9 financial skills and 5 compliance security measures. The solution compresses document processing and draft creation from days to hours, reducing financial modeling from 5-7 person-days to 0.5-1 person-days and deep research from 10-20 days to approximately 2 days. Dozens of financial institutions have already adopted Kimi products for business innovation, covering top securities firms, commercial banks, ventu
Zhongke Wenge releases SciencePro, the first publicly accessible AI research platform. Trained on 90PB of scientific data and 170 million papers, covering 32 million research projects and 700,000 industry reports, it offers project review, literature overview generation (compressing two weeks of work into minutes), and full research workflow assistance, integrating 2,000+ tools and 13 expert agents. Unlike general LLMs that merely call tools, SciencePro achieves fundamental scientific discovery
Keli Xinxu completes nearly 50 million yuan seed round led by Innovo科创 Fund. Funds will support AI4S foundation model iteration and automated experimental platform deployment. The team is building multimodal scientific data understanding foundation models covering DNA, RNA and proteins, while developing Agentic-VLA models for robotic wet lab experiments, aiming to achieve a cognitive-design-experiment-feedback-relearning dry-wet loop and construct an AI Scientist system capable of understanding
Anthropic released a progress report on AI R&D automation. As of August, Claude now leads 26% of AI R&D work, surging from under 1% in February. Approximately 30,000 AI agents operate simultaneously, generating over 1 billion decisions in August alone. This signals AI's transition from a supporting tool to a primary force in research and development.
Hujing Entertainment Group launches the "WhaleSharp AI Creation Competition" with a total prize pool of 10 million yuan and a top prize of 2 million yuan across 265 award slots. China's first AI video creation competition at this scale requires entries to be at least 10 minutes long with complete original stories, aiming to discover AI creators skilled in storytelling and narrative. Registration deadline: November 22.
YanRong's F9000X achieved 544GiB/s aggregate bandwidth in MLPerf Storage v3.0 testing, ranking first in both read and write performance for Checkpoint scenarios. The company has built a full-stack AI storage solution covering both training and inference, improving GPU utilization with hardware costs approximately 35% below market average.
Periodic Labs, founded by former OpenAI VP Liam Fedus, releases the Neon model, achieving 55.3% success rate on X-ray diffraction analysis benchmarks using only 1,300 H200 GPUs, outperforming GPT-6 Astra (requiring 100,000+ GPUs), demonstrating that real-world physics lab data is a new path for AI scaling.
Ziditaichu releases ZDTaichu5.0-9B, a general multimodal model that achieves 8 out of 9 first-place finishes in spatial understanding benchmarks at its parameter scale, outperforming models like Qwen3.5-9B while maintaining top-tier general multimodal capabilities, with full data pipeline open-sourced.
China Telecom's enterprise general-purpose Agent product, TeleAgent, ranked third in IDC's first-ever "China Enterprise General-Purpose Agent Product Technology Assessment." It scored 3.49 on routine tasks and 3.36 on complex tasks, earning full marks on task performance. With nearly 1.2 million users and a context window exceeding 400K, it achieves approximately 40% reduction in inference costs with 3-5 second response times for simple tasks. The team of over 1,200 people, with over 90% from so
TypeSafe AI releases Jev, a large language model specialized in high-frequency decision-making without content generation capabilities, serving purely as a classifier. Founder Diogo Almeida previously contributed to OpenAI's RLHF and InstructGPT development. Jev achieves 380ms average response time, 20-200x faster than traditional LLMs, at just $0.042 per million tokens, using RLCD training for probability calibration. Testing shows second-highest accuracy, fastest speed, and significant cost ad
Google DeepMind establishes the DeepMind Institute (DMI), led by Hassabis, Legg and Manyika. The institute focuses on AGI's impact on economy and society, bringing together researchers from AI, economics, philosophy and policy fields to jointly explore values, governance and what it means to be human in the AGI era.
Sharpa Robotics releases World Synesthesia Model at CoRL 2026, unifying visual geometry, tactile contact, proprioception and action history into a world model state for robust in-hand manipulation on 22-DoF dexterous hands. The model enables zero-shot generalization across 49 objects, with rotation capability improving nearly 3x after pre-training, drop rates reduced from 6% to 0.3%, and stable continuous rotation exceeding 1 minute.
OPPO unveiled ColorOS 17 at ODC 2026, with Jiang Yuchen outlining the AI phone strategy: self-developed on-device model as core, memory as foundation, and AIOS as execution layer. The company believes LLM gaps are narrowing while phone differentiation is just beginning. ColorOS 17 launches proactive features and memory background, boosting same-spec device background preservation by 55.6%, alongside new hardware Heart Sphere as an AI personal aide.
Harbin Institute of Technology (Shenzhen) and Pengcheng Laboratory release PolaFormer++, introducing a Polarity-aware Channel-wise Spiky (PaCS) feature mapping. This technique addresses negative value loss and insufficient attention sharpness in linear attention, enabling more precise differentiation between important and secondary information. It achieves 82.8% ImageNet accuracy with 5.25x speedup at 200K sequence length, and is open-sourced.
Researchers from Beihang University, CUHK and NUS release OmniHarness, enabling visual AI to accumulate verified workflow strategies through autonomous practice and result checking. Achieves 95% on ComfyBench creative tasks, surpassing the best baseline by 27.5 percentage points. Experience transfers to other systems like ComfyAgent and ComfyMind with improvements ranging from 10 to 24.5 percentage points.
OpenAI President Greg Brockman confirms significant progress on a second Millennium Prize Problem following the Navier-Stokes breakthrough. Solving P vs NP would destabilize modern cryptography. In the prior Navier-Stokes proof, OpenAI deployed approximately 10,000 AI agents continuously for 88 hours, sending 2.7 million messages and consuming 130 billion output tokens.
HiDream.ai releases HiDream V1, a native full-modality video generation model ranking in the global top 4 for image-to-video capability. The model uses an intent-first-then-generate approach, supporting 5-20 seconds adaptive duration and 1080p high-fidelity output. Tests demonstrate its ability to understand real-world physics, generating complex motion scenes like NBA buzzer-beaters and 100m sprints with smooth actions and synchronized audio-visual elements. This marks video generation's evolut