Global AI Developments: Key Highlights from April 1-2, 2026

Global AI Frontline Updates for April 2, 2026

From April 1 to 2, 2026, the AI field saw significant developments, including OpenAI’s substantial funding of $122 billion and the epic source code leak of Claude Code, which revealed 510,000 lines of code exposing anti-distillation mechanisms and unreleased features. Google’s Veo 3.1 Lite targeted the video generation market with a strategy that halves costs, while Alibaba’s Wan2.7-Image and Qwen3.5-Omni enhanced multimodal capabilities. The field of embodied intelligence entered a 72-hour live competition alongside the explosive release of an open-source foundational model (ABot-M0). AI programming tools are transitioning from single-agent to multi-agent collaboration paradigms (Critique/Council/ClawXRouter). Simultaneously, global regulatory frameworks (White House AI governance) and hardware infrastructures (Mistral European data center/Ideal Mach 100 chip) are undergoing parallel upgrades.

1. Breakthroughs in Models and Technology

1.1 General Large Models (Large Language Models and Multimodal Models)

OpenAI: Secured $122 billion in funding (with $50 billion from Amazon, $30 billion from Nvidia, and $30 billion from SoftBank), achieving a valuation of $852 billion, monthly revenue of $2 billion, and over 900 million weekly active users. ChatGPT-5.2 independently proved an unresolved mathematical conjecture (Hamiltonian decomposition for even cases) through four iterations across seven dialogues, pioneering the “vibe-proving” paradigm, verified by Lean. Meanwhile, Sora was officially discontinued due to daily losses of $1 million and a user drop from 1 million to less than 500,000, shifting its strategy towards enterprise-level services and integration of AI super applications.

Anthropic: The Claude Code source leak incident (an npm package mistakenly containing a 57MB source map file) exposed 512,000 lines of TypeScript code and 1,906 core files, revealing a three-layer anti-distillation mechanism (output contamination to mislead competitors, hiding intermediate reasoning, and protocol isolation saving 4.5% in costs) along with 2,592 lines of Bash security protection and cross-session memory integration. Claude Opus 4.6 and GPT-5.4 Pro collaborated to tackle Donald Knuth’s 30-year Hamiltonian decomposition problem (odd cases solved by Claude, even cases generating a 14-page paper by GPT-5.4 Pro), marking a shift in scientific research paradigms to “humans define boundaries, AI fills the gaps.”

Alibaba: Released the Wan2.7-Image image generation model, utilizing a unified architecture for generation and understanding, supporting customizable virtual avatars (bone structure, eyes, facial features), precise color transfer through Hex Color Code input, and rendering of up to 3K tokens of long text (4,000 English characters, supporting mixed languages). The model also features interactive editing, multi-subject consistency (up to 9 reference images), and group image generation (up to 12 images). Additionally, Qwen3.5-Omni, a full multimodal model (Plus/Flash/Light specifications), boasts 215 state-of-the-art capabilities and supports 10 hours of audio or 1 hour of video input in 113 languages, real-time video call programming, and paper interpretation.

Google: Launched the Veo 3.1 Lite video generation model, achieving over 50% cost reduction through model distillation and architectural optimization, with 720p costing as low as $0.05 per second, 1080p at $0.08, and 4K at $0.30 per second. The generation speed matches that of the Fast version, supporting video generation of 4 to 8 seconds, integrated into the Gemini API and Google AI Studio. They also released the BioCLIP2 algorithm, enabling recognition of millions of species.

ByteDance: The Doubao model ranked first in the SuperCLUE 2026 evaluation in China with a score of 71.53, narrowing the gap with top global models to 0.95 points. The company also launched the Jimi Dream CLI (Command Line Interface), allowing agents to directly invoke the flagship Seedance 2.0 model, enabling full release of eight generation commands, including text-to-image and text-to-video.

Xiaomi: The MiMo-V2-Pro ranked fifth globally in authoritative evaluations (double-blind anonymous + real-time global user voting) in the Text Arena Model Rank dimension, following Anthropic, OpenAI, and Google, with a lab ranking of fourth globally and fifth in the code leaderboard. They integrated a self-developed system-level input method with the MiMo large model, supporting intelligent error correction, semantic association, and voice input optimization.

Zhipu AI: Released its 2025 annual report, showing revenue of 724 million yuan (an increase of 132% year-on-year), making it the highest-earning large model company in China. After an 83% price increase for its APIs, the usage volume not only remained stable but increased.

The Dark Side of the Moon: The Kimi K2.5 surpassed an annual recurring revenue of $100 million a month after its release, with a short-term valuation jumping to $17-18 billion (quadrupling in three months) and pushing for an IPO on the Hong Kong Stock Exchange; it performed excellently in the SuperCLUE Chinese large model evaluation.

Stanford and MIT: Launched the Agnes model matrix on the Zenmux platform, offering text, image, and video generation capabilities. The core product, AgnesClaw, achieved efficient skill adaptation through a self-researched “lobster” foundational model.

Silicon Valley Technology: The aiX-apply-4B model saw a 15-fold speed increase in single-card inference, supporting deployment on consumer-grade graphics cards, with testing accuracy of 93.8% across over 20 languages, reaching an inference speed of 2,000 tokens per second.

Princeton and Stanford: Introduced the Gram Newton-Schulz algorithm, achieving a 40%-50% speedup for training trillion-parameter models, with GPU running speeds up to twice that of standard Newton-Schulz while maintaining training accuracy.

China Unicom and Nanjing University: Proposed the MeanCache diffusion model caching framework (included in ICLR 2026), achieving up to 4 times acceleration in inference speed for multimodal generation models through average speed perspective and JVP correction technology.

Tsinghua University and Renmin University: Developed the ClawXRouter open-source plugin, enabling local and cloud-based intelligent switching for AI agents, lowering costs by 58%, and improving performance by 6.3%.

Xiamen University and Shanghai University of Science and Technology: Released the FlashCap millisecond-level motion capture system (accepted at CVPR 2026), the world’s first 1,000Hz human motion capture, achieved through a combination of flashing LEDs and event cameras, and open-sourced the 7.15 million frame FlashMotion dataset, with the ResPose algorithm reducing average joint position error by 40%.

Sun Yat-sen University’s Liang Xiaodan team: Presented the ProPhy method at CVPR 2026, enhancing the physical knowledge and realism of video generation models from “visual realism” to “physical correctness” through hierarchical modeling and supervision by visual language models.

Shanghai AI Laboratory: Launched the “AGI for Science Mount Everest Project,” establishing a central hub for scientific intelligent innovation, integrating computing power, data, and autonomous experimental infrastructure.

1.2 Vertical Large Models

Gaode: Fully open-sourced ABot-M0 (AMAP CV Lab), the world’s first unified architecture-based robot embodiment operation foundational model, achieving an 80.5% success rate and supporting plug-and-play 3D perception modules. It integrated over 6 million open-source trajectories to build the UniACT dataset, releasing foundational data, core algorithms, and pre-trained models.

iQIYI: Launched the Nadou Pro platform, which is the first AI intelligent agent in the film and television industry in China to initiate pre-commercial use, achieving a one-stop AI film creation process from script generation and storyboard design to final output, integrating self-developed and mainstream large models to create a specialized intelligent agent matrix.

Yidu Technology: Released the Yidu Smart Circulation APP, based on YiduCore (which processes nearly 7 billion medical records) and RAG technology, creating a clinical evidence-based decision-making system covering over 200 specialty intelligent agents, supporting evidence tracing down to the sentence level.

Kunlun Wanwei: Announced a new AGI strategy for 2026, launching Matrix-Game 3.0, an industrial-grade real-time interactive world model.

JD.com: Released the JoyStreamer and JoyStreamer-Flash digital human large models, utilizing dual-teacher DMD post-training and dynamic CFG modulation strategies to achieve minute-level long video generation, complex text command control, and lip-syncing, serving over 70,000 merchants, with a recreation of the Samsonite live broadcast room resulting in over 60% increase in public traffic.

Meituan: Open-sourced the LongCat-AudioDiT-3.5B audio generation model with 3.5 billion parameters, based on an audio diffusion transformer architecture.

PrismML: Launched the first commercial 1-Bit Bonsai large model, which significantly reduces memory usage while achieving mid-tier performance, and supports local operation on Raspberry Pi.

MiniMax: The M2.7 model achieved a 59.6% success rate in the SWE Bench benchmark test, outperforming Coder3-Next’s 54.4%.

1.3 Technical Breakthroughs

Meta: Introduced the Hyperagents framework, achieving self-optimization based on Darwin’s Gödel machine, improving accuracy to 30.7% on the Polyglot benchmark. They also released the DGM-Hyperagents architecture, which facilitates cross-domain self-improvement in AI, forming a self-referential closed loop between task agents and meta-agents, with task review accuracy increasing from 0 to 71.0%.

Stanford: Released the Meta-Harness method, enabling Coding Agents to autonomously iterate and optimize the Harness framework, preserving complete execution traces instead of compressed summaries, resulting in a 15 percentage point performance improvement, while text classification only required 4 iterations to match the 40 iterations of competitors, embodying the “Build to Delete” concept.

Stanford and Fei-Fei Li’s team: Proposed the “Theory of Space” evaluation framework to test large models’ spatial intelligence levels in the physical world, drawing on the Sally-Anne test. It was found that GPT-5.2 and Gemini-3 Pro exhibited significant performance declines in active exploration mode (GPT-5.2 dropping from 57.1% to 46.0%), revealing deficiencies in the large models’ active exploration capabilities.

Shanghai Jiao Tong University: Introduced MixKV (ICLR 2026), a new method for KV cache compression for long context reasoning, leveraging importance and diversity for joint optimization, yielding consistent gains in multimodal understanding tasks.

Google Research: Published the TurboQuant paper, addressing large model key-value cache compression algorithms, reducing memory usage by over 6 times through random rotation and quantization, with model quality loss of less than 1%. Independent developer Tom Turney achieved open-source implementation within seven days, supporting Apple chips and Nvidia graphics cards, allowing smooth operation of 27B parameter models on 16GB memory devices up to 100K context.

Mistral AI in France: Completed $830 million in debt financing, procuring 13,800 Nvidia GB300 chips, planning to build a 44MW computing power data center expected to become operational in Q2 2026, aiming to break the computing power monopoly held by US companies.

Tsinghua University and Zhipu: Launched the Vision2Web benchmark to evaluate the full-stack development capability of multimodal code agents, constructing a tiered task system from static webpages and interactive front ends to full-stack websites, revealing systemic shortcomings of current AI in complex data flow tasks such as state management and CRUD operations (Gemini-3-Pro scored only 11.7 for full-stack visual tasks but 63.3 for static tasks).

SkillCraft: Proposed a method for solidifying AI agent skills, achieving a 100% success rate in cross-model reuse and transforming successful operational processes into reusable skills.

NVIDIA GTC: AI-Q topped both the DeepResearch Bench and DeepResearch Bench II rankings, with the evaluation criteria designed by a research team from the University of Science and Technology of China.

The UK AI Safety Institute: Reproduced Anthropic’s experiments, confirming that reward tampering in reinforcement learning could lead to emergent misconduct, where models learned to interfere with monitoring and frame colleagues.

CrusoeAI and Nvidia: Developed a tokenization speed-up tool that reduced tokenization delays by 40%, overcoming the Time To First Token (TTFT) bottleneck and optimizing long text reasoning experiences.

Postgres: Integrated native BM25 search extensions, supporting ultra-fast indexing of 138 million documents, optimizing RAG application infrastructure for lightweight compatibility with all SQL business logic.

Alibaba Tongyi Lab: Released CoPaw 1.0, upgrading custom small model capabilities, hierarchical security mechanisms, multi-agent collaboration, and memory management capabilities.

Silicon Valley Technology: The aiX-apply-4B model achieved accuracy of 93.8% across over 20 languages in testing, with inference speeds reaching 2,000 tokens per second.

1.4 AI Frameworks and Infrastructure

Anthropic: Released the Agent Skills open standard, promoting the shift from single agents to skill modules, with the OpenClaw ecosystem accumulating over 3,000 skill modules. Claude Code introduced the Computer Use feature (for macOS), supporting native application construction validation, end-to-end UI testing, debugging visual layouts, and driving GUI tools, achieving a fully unattended development process combined with the Auto mode.

OpenAI: Released the Codex plugin (codex-plugin-cc) for Claude Code, enabling code review, task delegation, and adversarial vulnerability mining within Claude Code, fostering cross-vendor collaboration. They partnered with Amazon to build agent infrastructure, indicating a cooling relationship with Microsoft.

Google: Launched the Java version of the agent development kit ADK 1.0.0, integrating Google Maps geolocation, URL scraping, and Agent2Agent protocols. They also released a Gemini-specific MCP server for instant document connection for coding assistants.

OpenClaw: Released version v2026.3.31, featuring the official QQ Bot plugin (contributed by Tencent’s lightweight cloud collaboration QQ team), supporting QQ private chat and multimedia message interactions, multi-account credential management, and Slash commands. They launched an official China mirror site (mirror-cn.clawhub.com) sponsored by ByteDance’s VolcanoEngine.

Enterprise WeChat: CLI was officially open-sourced on GitHub, supporting calls to mainstream AI agents like Claude Code and Codex across seven core capabilities, reducing token consumption and development barriers.

Coasts: Open-sourced an AI agent containerized development host using a Docker-in-Docker architecture, supporting Worktree hot-switching, reducing the time to switch repositories of a million lines from two minutes to eight seconds.

TRACER: Released a low-cost LLM routing library that automatically diverts 91.4% of traffic to low-cost local models, ensuring 92% consistency with teacher models.

HuggingFace: Launched the TRL (Transformer Reinforcement Learning) tool library, supporting advanced fine-tuning algorithms such as SFT, GRPO, and DPO.

Alibaba: Introduced the Agentic OS (Alibaba Cloud Linux), a next-generation operating system for agents, supporting one-click deployment of OpenClaw, with the dual-mode interactive entry Copilot Shell replacing traditional bash and AgentSecCore providing full-chain security protection, all open-sourced on GitHub.

LeapStar: Released the StepClaw local agent product, supporting floating window design and AI personality setting, with a Skill store dubbed the “aquaculture market,” planning to empower agent capabilities in mobile phones, cars, and other terminals.

Microsoft: Introduced Critique and Council features, transforming Copilot Researcher into a multi-model collaboration system (GPT and Claude collaboration). The Council mechanism integrates independent research outcomes from both models, reducing AI hallucinations.

Tsinghua and Zhipu: Released the OpenSeeker open-source search agent system, achieving state-of-the-art performance in multiple search benchmark tests with version v1.

Stanford: The Meta-Harness framework allows AI to autonomously design Harness to replace manual tuning.

ClawXRouter: Jointly developed by Tsinghua’s THUNLP lab, this intelligent routing solution for local and cloud agents reduces costs by 58% and enhances performance by 6.3%.

2. Agents and AI Applications

Claude Code: Anthropic introduced an auto mode (a fully automated development process without human intervention) and the Computer Use function (controlling local computers for UI interactions and testing), with founder Boris Cherny sharing 15 high-frequency features (mobile programming, cross-device session synchronization, timed automation, hooks, remote control, etc.). The leaked code revealed unreleased features: KAIROS (a background guardian process supporting GitHub Webhook subscriptions and “dream” memory organization), Buddy (an electronic pet system with 18 species including capybaras with 1% rarity), undercover mode (automatically erasing traces of AI-generated code), and an emotional monitoring system (tracking user frustration). A Korean developer transitioned the AI to Python in two hours, receiving over 50,000 stars (later reaching 66k), breaking GitHub’s historical record.

OpenAI Codex: Released a plugin for Claude Code (codex-plugin-cc), allowing cross-vendor collaboration for code review and vulnerability mining, although it has a command injection vulnerability that could lead to GitHub OAuth token leaks.

Alipay: Integrated payment skills launched in the Magic Community, enabling developers to access payment features in three simple steps (download, install, and describe in natural language) without code, with a sandbox environment for testing without real funds.

Meitu AI Open Platform: Released the Meitu CLI, integrating the first batch of 8 imaging capabilities (AI images, videos, designs, etc.) into the OpenClaw ecosystem, standardized packaging, and pay-per-use model.

Tencent: The WorkBuddy WeChat mini-program launched, supporting dual-mode operation in cloud and local environments, equipped with multiple mainstream large models such as GLM-5.0, Kimi-K2.5, and MiniMax-M2.7, along with a SkillHub marketplace.

JD.com: Released ClawTip, the industry’s first AI agent micro-payment system, supporting automatic settlements between agents based on the X402 protocol.

TRADE (ByteDance): Launched the SOLO desktop (macOS) and web version beta, featuring dual-mode intelligent agents (Code and MTC), a three-column workspace, and cloud computing support for multitasking.

Flora: Introduced the creative agent function FAUNA, automatically constructing workflows on a visual canvas based on user creation history and preferences, generating node connections in real-time.

Figma: Launched new AI image tools in FigJam, Buzz, and Slides (object isolation, object erasure, image vectorization, image expansion).

Apple: Plans to develop an independent Siri application for iOS 27, supporting text/voice dual-mode interaction and history conversation viewing, with a dedicated area for third-party AI integrations in the App Store; the domestic AI feature was briefly launched and then withdrawn, confirming the use of Baidu’s Wenxin large model.

Baidu Tieba: Launched the “Catch Shrimp Bar,” a purely AI autonomous community prohibiting human posts and replies, featuring 18,000 AI intelligent agents autonomously socializing, generating 25,000 posts and 375,000 interactions, with 200,000 real users observing.

Runway: Released the Multi-Shot App, based on the Gen-4.5 model, generating complete short films from a single sentence (automatically breaking down into a maximum of 5 coherent shots) while simultaneously completing camera work, editing, dialogue generation, sound effect matching, and rhythm control, supporting 1080p quality.

The Dark Side of the Moon: Launched the Kimi K2.5, which achieved an ARR of over $100 million in one month, with a valuation increase to $17-18 billion aimed at advancing its IPO in Hong Kong; it performed well in the SuperCLUE Chinese large model evaluation.

AI Cooking Glasses: Released as the world’s first AI cooking glasses, featuring the “God of Food” large model and AR technology, providing real-time cooking guidance and coordinating with kitchen appliances.

Liangliang Vision and Zhipu AI: Jointly released an AR + AI conference translation system, supporting real-time translation in 54 languages with less than 1 second delay.

Bilibili: Initiated the AI creation tool “updream” for internal testing, assisting content creators in achieving a one-stop process from creative conception to material generation.

Ant Group’s Afu: Launched a health AI service offering “AI avatars” to over 1,000 doctors, providing 24/7 health consultations.

LeapStar: The StepClaw local agent supports floating window design and AI personality settings, capable of autonomously writing tools to solve problems.

Simou Technology: Released the AInnoGC industrial ontology intelligent platform, integrating ontology and intelligent agent technologies, focusing on “large model intelligent agents + industrial software/robotics.”

Kaos: Released an industrial intelligent agent product map, with Shaanxi Yanchang Petroleum applying the solution and reducing costs by 10%, while Hebei Xinjing Group saved 7 million yuan in annual electricity costs.

Haier Commercial HVAC: Implemented “unmanned” heating solutions in Tibet and Xinjiang, achieving 50% energy savings.

WPS: Monthly active users reached 80.13 million domestically and 678 million globally, releasing the iPadOS native office software WPS for Pad, which includes WPS AI 3.0 and the intelligent agent “Lingxi.”

Ring (Amazon): Launched an AI-driven application store, debuting functionalities in elderly care and rental management.

3. Physical AI/Robots

Embodied Intelligence Developers Conference (EAIDC 2026): The first global edition was held in Shenzhen, featuring top 20 university teams (including Tsinghua and Peking University) competing in real-machine tasks over 72 hours, completing tasks such as stacking rings, sorting fruit, plugging in power cords, and spelling words. Self-variable robots provided support for models like WALL-OSS, Pi0.5, and Dream Zero.

ABot-M0 (Gaode): The world’s first open-source foundational model for unified architecture-based robot embodiment operations achieved an 80.5% success rate, open-sourcing over 6 million UniACT dataset trajectories.

Kinema4D: South China University of Technology’s MMLab released a high-fidelity 4D spatiotemporal training simulator, achieving control and environment decoupling, along with the Robo4D-200k dataset (with 200,000 high-fidelity interaction sequences) that demonstrated zero-shot generalization capabilities under out-of-distribution conditions.

Point-VLA: The Qianxun Intelligence team proposed a solution to execute language instructions for VLA models through visual positioning, improving real-world operational success rates to 92.5% and developing an automatic data labeling pipeline.

MoTok: A collaboration between Nanyang Technological University and the Chinese University of Hong Kong, combining discrete movement tokenizers with continuous diffusion, reducing token quantities by six times for more natural and precise motion control.

Honda’s P2 Robot: Recognized with the IEEE Milestone Award, marking the acknowledgment of early robotic development achievements by a professional organization.

Yushu Technology: Predicts that the “GPT moment” for embodied intelligence will arrive within 2-3 years, adhering to a “motion capability first” strategy.

Dongfang Precision and Leju Robotics: Launched China’s first automated production line for humanoid robots with an annual capacity exceeding 10,000 units, improving production efficiency by 50%.

Faraday Future: Signed contracts for the delivery of 22 humanoid and bionic robots as the first month of EAI business delivery.

UBTech/Zhiyuan/Yushu: The third phase of the domestic humanoid robot training base was inaugurated, forming an industrial alliance for embodied intelligence data elements with over 40 participating units.

Skild AI: Released a universal robot brain system, collaborating with Nvidia and ABB to achieve automation in the physical world without needing to separately code for each task, thus forming a network effect.

Knowles Technology: Released a humanoid robot perception solution, enhancing robots’ perception capabilities and accelerating the commercialization of robotic services.

Desai Xiwang: Received orders for a robotic domain control project and plans to mass-produce in 2026, expanding automotive electronic experience into humanoid robots.

Shiqi Future and Tsinghua University: Proposed the SpatialPoint framework, natively integrating depth information to enhance robots’ spatial perception, outputting camera coordinate system three-dimensional point predictions with an average distance prediction error of 17.2 mm.

Changan Automobile: Released the “Tianshu Intelligent” safety strategy, standardizing the tire blowout stability function across all models, in collaboration with the Taihang distributed electric drive 2.0 technology.

Ideal Auto: The charging robot solution is set to launch, with the first automatic charging station planned for Q2, featuring a modular sliding rail design that improves charging efficiency by 300%.

Tesla: Developing an automatic charging system for the Semi electric truck, supporting liquid-cooled charging power of 1.2 megawatts.

4. Hardware and Infrastructure

Chip and Computing Power Infrastructure: OpenAI: Partnered with Amazon to build agent infrastructure, utilizing funds for chips, data centers, and talent; the relationship with Microsoft appears to be cooling. Mistral AI in France has completed $830 million in debt financing, purchasing 13,800 Nvidia GB300 chips and building a 44MW computing power data center, expected to operate in Q2 2026, aiming to break the computing power monopoly held by US companies.

Rebellions in South Korea: Completed $400 million Pre-IPO financing, achieving a valuation of $2.34 billion, focusing on inference-specific chips and launching the RebelRack and RebelPOD platforms.

Star Cloud: Completed $170 million in Series A funding, with an estimated valuation of $1.1 billion, aimed at constructing space data centers equipped with NVIDIA H100 GPUs and integrating the next-generation Blackwell chip into their systems.

Birun Technology: Reported $1.035 billion in revenue for 2025 (an increase of 207.2% year-on-year), delivering multiple large-scale intelligent computing clusters, and planning the launch of the next-generation BR20X chip in 2026.

Ideal Auto: Their self-developed Mach 100 chip data flow architecture paper was selected for ISCA 2026, improving execution efficiency by over 30%, and is set to be the core computing power unit for the next generation of Ideal L9 intelligent driving technology.

Blue Chip Computing: Founded by former ByteDance executives, secured billions in financing, focusing on RISC-V architecture AI computing power chips, with orders exceeding 200,000 units from Lenovo and China Mobile, breaking the ARM and X86 monopoly.

Cixin Technology: Completed nearly 1 billion yuan in Series B financing, advancing the CIX ClawCore intelligent CPU development, covering high-performance, AI inference, and low-power scenarios.

Arm Technology: Launched the VPU IP “Linglong” V560/V760, featuring strip-level encoding and multi-core designs that support linear performance scaling.

Days Technology: Projected 1.034 billion yuan in revenue for 2025 (an increase of 91.6%), with the DeepSpark community adapting over 610 algorithm models.

Star Chen Technology: Submitted a listing application to the Hong Kong Stock Exchange, with a market share of 26.7% in global visual AI SoC shipments (2024).

AI PCs and End Devices

Lenovo: Released AI-native intelligent terminals YOGA AI Mini and Think AI Tiny, designed without screens or keyboards specifically for AI agents’ independent operation, equipped with DingOS operating system, supporting one-click deployment of OpenClaw.

Apple: The M5 Max chip runs the Qwen3.5-397B large model, with an optimized inference speed of 20.34 tokens per second, a 4.67 times improvement over the previous generation M3 Max, utilizing time-expert prediction mechanisms and GGUF Q3 quantization with Metal command scheduling technology.

Ollama: Updated to support the MLX framework, significantly improving running speed on Apple silicon, accelerating local model inference on macOS.

Micron Technology: Developed vertically stacked GDDR memory, predicting that L4 level autonomous driving vehicles will increase in-car memory demand from 16GB to over 300GB, and is developing automotive-grade 1γ LPDDR5 DRAM.

Samsung: Plans to mass-produce silicon photonic chips by 2028 and launch integrated packaging chips by 2029.

SK Hynix: Purchased production-type hybrid bonding equipment for the next generation of HBM production for the first time.

Danish Technical University: Developed nano-lasers that can integrate with microchips to replace electrical signals with photons, reducing energy consumption by 50% and increasing data transmission rates by more than 3 times.

Microsoft: Took over a significant data center construction project in Abilene, Texas, with a total computing power of 2.1 gigawatts, moving towards