TypeSafe AI's Jev sees valuation soar from $200M to $10B in 9 days, with $1B+ funding talks underway. Jev specializes in rapid decision-making, achieving 194x speed and 445x cost advantages over traditional language models, designed specifically for software and Agent applications.
Agent Space is a cloud-based agent workspace aggregating top tools like Claude Code, CodeX, and OpenCode, supporting local conversation import for cloud execution. A $10 subscription provides access to latest models including Opus 5.5 and GPT-6 Astra, with quota transferable to local agents. Friend referral offers 15% cashback plus 10% first-purchase bonus, lowering barriers to accessing advanced AI models.
OpenAI is testing a $500/month ChatGPT Pro Max subscription, 2.5x the current $200 Pro tier, with core卖点 being compute priority rather than usage volume. The trigger was GPT-6 Astra's massive compute consumption, leading to last week's suspension of new Pro subscriptions. This will surpass Claude Max ($250) as the new ceiling, with more details expected at next week's DevDay.
Meta unveiled Muse Charm, a handheld AI device designed by former Apple design chief Alan Dye, featuring a 2-inch OLED screen, 5G modem, and fingerprint wake sensor. The device gives the AI assistant a portable physical form, transforming it into a "digital pet" attachable to keychains. Muse climbed to the top of the US iOS free app chart within two weeks of its September 8 launch, reaching 730,000 downloads in five days and surpassing 2.5 million total—outperforming ChatGPT's launch metrics. Or
OpenAI launches a major update for ChatGPT Voice, adding voice-operated task capabilities. Users can now simply speak to have AI check emails, schedule meetings, create PowerPoints, build websites, and recover duplicate charges—eliminating manual operations entirely. The full GPT-6 family (Sol/Luna/Astra) powers the system with prices halved. The new version rolls out globally today, with Plus members getting 3 hours of voice access daily.
iFlytek releases speech recognition model Spark-ASR-2.0, trained on domestically-produced computing power, supporting 202 dialects without manual switching. The model features contextual error correction and text normalization, automatically converting speech into fluent, standardized text to reduce proofreading workload. Inference cost increases by only 10% compared to the previous generation, now available on iFlytek Input.
Meta launches personal AI agent Muse, achieving approximately 2.8 million global installations in 12 days, with downloads approaching 360,000 on September 20 alone. Its core design was exposed as originating from open-source project OpenClaw, including the 'soul file' SOUL.md. Muse cloudifies OpenClaw's philosophy, reducing usage barriers through secure cloud VMs and isolation, providing AI assistant services for ordinary users. OpenClaw's author Peter Steinberger joined OpenAI in February this
Former Doubao PC head Qi Junyuan launches Agent OS Today, featuring Memory and Proactive capabilities that remember users' contacts, projects, preferences and todos, generate personalized daily briefs, sync relevant updates, connect with calendars, email, notes and third-party services like GitHub and Figma, powered by a 24-hour cloud computer for complex tasks, officially released today.
Zhang Xiao Hu of GSR Ventures suggests this year may be the last for foundation models. As AI capabilities homogenize, user context emerges as the core competitive asset. Open-source project Orbital introduces 'Context is yours' concept, allowing users to accumulate context into a portable 'memory USB drive.' Combined with TokenDance's 70+ model access and transparent pricing, users gain full control over their context assets, preventing vendor lock-in.
SoftBank plans to issue approximately $11 billion in high-yield bonds with yields of 9%-10% and maturities of 3.5 to 7.5 years, with proceeds designated for the October payment due to OpenAI. If completed, this would become the largest non-financial corporate bond issuance in the Asia-Pacific and Japan region. SoftBank has raised approximately $37 billion this year, with cumulative investment in OpenAI approaching $65 billion for an approximately 13% stake. Masayoshi Son opts for long-term bonds
Eight Sleep launches smart mattress cover Pod, using AI to automatically control temperature and adjust firmness, improving sleep performance by 30% and predicting disease risk with nearly 90% accuracy. Backed by investors including Sequoia China, the company is valued at $1.5 billion and turned profitable last year. After entering China in April 2026, it became the company's fastest-growing market.
Scinetics completes ~50M RMB seed round led by Inno科创 Fund, to develop AI4S foundation models and automated experiment platforms. Founder Zhang Zaixi, a USTC PhD, serves as HKUST Assistant Professor. STELLA scientific AI agent system serves over 1,000 researchers, with international users from Stanford, MIT and other institutions.
Seven universities jointly open-source OpenWAM, a 'world simulator' for robotics. Through modular infrastructure and large-scale pretraining validation, it achieves top-tier results across eight simulation benchmarks and three real robot platforms, providing a reproducible foundation for world-action model research.
At the 2026 Yunqi Conference, Banma Intelligence unveiled AutoOmni 2.0-23B-A3B edge large language model, featuring a sparse architecture with 23B total parameters and 3B activated. The model achieves near-cloud-scale performance in cabin scenarios, with 99%+ quantization fidelity, 50% memory reduction, and 5-6x inference speedup, backed by a three-layer security architecture and ASIL-B automotive certification. The company is expanding this edge-cloud collaborative solution from cockpit to auto
Anthropic releases Claude Opus 5.5, demonstrating strong autonomy: given a complete task, it autonomously plans and executes. Users generated 100 HTML pages in different styles in one go. Performance improved significantly: reviewing and fixing 200K lines of code took under 3 hours (vs 20+ hours for Opus 5); rewriting HAProxy from C to Rust took 9.5 hours. It also creates JavaScript animations and music, with substantially improved writing style for more natural output.
GLM-5.3-Flash combines Kimi's KDA and DeepSeek's KPool-DSA across its 45-layer architecture. MiniMax's反复 iterations from Linear to Full to Sparse reveal that as specific techniques become verifiable, absorbable, and combinable, attention architectures are shifting from seeking a single optimal mechanism toward letting different mechanisms handle different tasks, making the "technology route" label increasingly obsolete.
In September 2026, the AI Native Development Practice Forum was held at the Cloud Computing Conference. Six experts highlighted that while code generation accelerates, the real bottleneck has shifted to the complete delivery chain: coding occupies only 20%-30% of delivery time, with key challenges being requirement clarity, context continuity, environment reproducibility, result verification, and production governability. Ant Group reduced badcase fix time to one-quarter, while Fliggy shortened
Huawei's Ascend 950 SuperNode launches as the industry's largest-scale commercial supernode. The Atlas 950 SuperPoD liquid-cooled supernode features 1024 NPUs with Lingqu high-speed interconnect technology breaking storage and communication walls, delivering 1.5-1.7x training efficiency improvement and 2-3x inference performance with sub-10ms latency. The Atlas 850E air-cooled supernode enables 10ms-level inference with 2.5x per-card throughput improvement, deployable directly in standard air-co
Knowin releases generative learning framework GLOW and robot butler X1, achieving 'one-demonstration learning': robots learn to execute tasks after watching human demonstrations. GLOW achieves 62.2% success rate on RoboDojo benchmark, outperforming GPT-6's 22.2%; 86.7% on LIBERO-Pro vs GPT-6 Astra's 58%; and 65.62 score ranking first among 20 embodied models on Embodied Arena. The system uses native autoregressive model unifying vision, language and action, combined with Harness closed-loop feed
Mifeng Technology launches "MifengPai", the world's first full-category physical AI data crowdsourcing platform. Users download the app, apply for MEgo devices, and earn money by collecting data part-time. The platform has deployed 20,000 device sets, accumulated 1 million hours of data across 22 major scenario categories. The team states that本体-free data is the solution to embodied AI training bottlenecks, with the socialized collection model enabling "everyone to become a robot training specia
DeepSeek and Tsinghua University jointly released a paper on DSec sandbox platform (co-authored by Liang Wenfeng) for training AI agents. The platform enables agents to build their own training environments within sandboxes, achieving an RSI loop where agents create environments, environments train agents, and stronger agents build better environments. Performance reaches 3 million sandboxes per day, 380,000+ peak concurrency, and 5,000+ creations per second.
Huawei unveiled Kunpeng Hyper Node upgrade, with dual flagship solutions supporting up to 4096-node networking and building a 100TB-level unified memory pool. The 100GB-level bandwidth and nanosecond-level latency define new hyper node standards. After seven years of ecosystem building: 300+ server models, 28,000+ industry solutions, 4.16M global developers. Trading architecture for compute power, Kunpeng Hyper Node enables thousands of servers to work in concert like a single supercomputer, pro
Researchers from Zhejiang University, Tsinghua University and NUS introduce TurboT2VA, a three-stage curriculum distillation approach that compresses a 40-step teacher model into a 4-step student. It achieves 20.1× speedup at standard resolution and 54.67× speedup at high resolution, reducing generator time from 318.74s to 5.83s while maintaining audio-visual alignment and generation quality.
Qwen releases "Personal Agent" upgrade, enabling the model to help users complete complex tasks like ordering, shopping, and booking transportation. By integrating authorized personal data like health and fitness information, and leveraging tools inherited from Quark like cloud storage and question banks, it delivers personalized services. Data shows 5x growth in complex tasks with 5-8 conversation turns. Qwen aims to evolve from ChatBot to Personal Agent, becoming a true personal AI assistant.
The embodied AI industry is grappling with data scarcity, with companies globally establishing data flywheels covering seven stages: collection, storage, cleaning, annotation, quality inspection, augmentation, and training evaluation. Alibaba Cloud provides full-stack engineering capabilities, compressing some enterprises' training feedback cycles from weekly to daily. Qiange Intelligence achieves 80%-90% AI-driven annotation, T-Head Zhenwu 810E chips have shipped over 10,000 units, and Mocke Ro
METASTONE releases Meta-Infer, an enterprise-grade high-performance LLM inference engine, achieving a 6.87x throughput improvement (from 1,932 to 13,274 tok/s) for DeepSeek-V4.1-Flash on 8-card PCIe infrastructure through pure software optimization including kernel completion, operator fusion, communication refactoring, and cache reuse. The engine supports 1M context length and has demonstrated 1.92x throughput gains for GLM5.3 and 2.48x video generation speedup for MiniMax H3, proving that soft
The 2026 PT Expo marks a new phase of AI implementation: the theme shifted from Post & Telecommunications to People & Technology. Telecom operators are exploring Token economics, with TeleAgent exceeding 750,000 registered users and 200 billion daily Token consumption. Galaxy Universal's robots have surpassed 1,000 commercial deployments across approximately 50 cities and nearly 200 smart pharmacies. The PT Expo documents the industrial transformation from switches to AI, as AI moves from innova
Qualcomm unveiled two flagship SoCs: Snapdragon 8 Elite Extreme Gen 6 and Snapdragon 8 Elite Gen 6, manufactured using TSMC's 2nm process, marking the industry's first 5GHz smartphone CPU clock speed. The new chips are optimized for on-device AI agents, supporting up to 30B parameter MoE models with 40% power reduction, shifting the computing paradigm from mobile-centric to AI Agent-centric.
AIsphere releases PixVerse R2 real-time world model, topping X trending list on September 22. The model enables real-time, continuous interactive audio-video generation through Omni Causal AR and Real-Time Acceleration technologies. Users can control characters via WASD keys and modify world elements through natural language, achieving four core capabilities: long-range consistency, multimodal control, causal response, and state preservation. Users transition from viewers to participants as AI d
When AI independently completes tasks, young doctors lose surgical practice opportunities and new consultants miss the learning process from making mistakes. A Turkish study shows students using regular AI for practice scored 17% lower on independent tests. Scholar John Ruskin noted the industrial era "manufactures everything except people." The article argues that when the market stops paying for a capability doesn't mean individuals can afford to lose it—judgment and understanding must form th
Peking University's DCAI team releases DataFlex-RL, providing configurable components for data selection, reweighting, and domain mixing. The framework integrates with verl training流程, enabling researchers to quickly deploy data strategies and compare their effectiveness fairly. Across 591 experiments, GRPO training boosted Qwen2.5-7B-base by 7.76 points and Llama-3.1-8B-base by 10.25 points, ranking #2 on HuggingFace Paper of the Day.
DeepSeek releases the Harness desktop client with a complete GUI, supporting four workflow modes: Standard, PTC, Minimal, and Creative, with six built-in official plugins. This significantly lowers the usage barrier compared to the command-line version. Signature verification confirms it is officially from DeepSeek, though it has not been formally announced and currently does not support Linux.