Google has launched Nano Banana 2 Lite, an AI image model that can generate an image in 4 seconds. It costs $0.034 per 1000 images, emphasizing speed and batch processing capabilities, making it suitable for high-frequency content production scenarios. The model is available on Google AI Studio, Gemini API, and enterprise platforms, and will replace the original Nano Banana.
Anthropic has released Claude Sonnet 5, positioned as its most agent-capable Sonnet model. It supports planning and the autonomous execution of complex tasks using tools such as browsers and terminals. Its overall capabilities are close to Opus 4.8, but at a lower price. The model is available across all plans, becoming the default model for Free and Pro users, and supports Claude Code and Platform calls.
Global fintech platform Airwallex announced the completion of a $320 million Series H funding round, bringing its post-money valuation to $11 billion, a jump of approximately 40% from $8 billion in six months. The round was led by Addition, with participation from several international institutions. The company also launched two AI products: the AI-native financial platform T:0 and the smart consumer wallet Airi. The funds will be used for AI product development, global compliance efforts, and team expansion.
Huawei has officially open-sourced its openPangu-2.0-Flash model, boasting 92 billion parameters. Based on Ascend's native training and inference technology, it serves as an intelligent foundation for the Agent era. Model weights, basic inference code, and training/inference operators are already available on the open-source platform. The Pro version will be released in July, with more components to be launched in the second half of the year.
Tsinghua-affiliated startup Liqing Intelligence has completed a seed round of financing worth hundreds of millions of yuan, with participation from Shunwei Capital, Sequoia China, and Hillhouse Capital. The team was founded by Li Yiming, an assistant professor at Tsinghua University's School of AI, and has an average age of 23. The company rejects the "world model" label, positioning itself as a Physical AI infrastructure provider. It has developed its own data pipeline and differentiable physics engine, achieving a Real-to-Sim-Real closed loop, and can achieve equivalent training results with only 1% real-device data.
Tencent Video has launched WorkRally, an industrial-grade AI-powered animation production platform for 2D/3D animation and AI-simulated human-driven dramas, covering the entire chain from script analysis, storyboard generation, content production to asset management. The platform features a built-in S+ level film and animation skill library, supports expert-level agent generation and intelligent pipeline scheduling, and can achieve character consistency management and continuity error detection.
Meituan has officially open-sourced its trillion-parameter large-scale model, LongCat-2.0. With a total of 1.6T parameters and an average activation of 48B, it is the industry's first model to complete the entire training and inference process on a domestic computing cluster with 50,000 GPUs. The model natively supports 1MB ultra-long contexts and is optimized for Agentic Coding tasks. The preview version has already ranked among the top three in global OpenRouter usage.
On June 30th, the official OpenAI developer account announced on the X platform that it will collaborate with custom keyboard manufacturer Work Louder to launch a dedicated input device under the Codex brand, scheduled to go live on July 15th. The device is based on the Creator Micro 2 micro keyboard and features 12 customizable mechanical keys, an eight-way joystick, and a stepless rotary knob, optimized specifically for AI development scenarios.
X Square Robot completed four rounds of financing within two months, with its post-investment valuation exceeding 20 billion yuan, becoming the first embodied brain company in the Greater Bay Area to achieve a valuation of over 20 billion yuan. More than 30 investors participated, with internet giants Meituan, Alibaba, ByteDance, and Xiaomi leading each round. Xiaomi invested in three consecutive rounds, and 58 Group, Chery, China Mobile, Sequoia China, and others also joined the investment.
Alibaba's Tongyi Qianwen team has launched an independent AI voice input method, "Qianwen Input Method," emphasizing "speak freely, write beautifully." The product integrates with Qianwen's large-scale model, CosyVoice, enabling speech-to-text conversion, semantic understanding, automatic removal of interjections, and correction of slips of the tongue, transforming scattered spoken language into structured written text. Qianwen Input Method supports input speeds of up to 300 characters per minute, recognition of 9 dialects, and mixed Chinese and English speaking.
Shengshu Technology has launched Vidy, an AI companion app positioned as "AI Vibe Buddy," supporting text, voice, and real-time call interaction. The product features differentiated AI characters such as Humi and is equipped with a relationship growth memory system that accumulates understanding of user habits and emotions through conversations. Leveraging its self-developed U-ViT architecture and Vidu video model, Vidy plans to achieve real-time generation of AI avatar expressions and movements, pushing companionship from text and voice to the level of face-to-face video calls.
Gaode Maps is internally testing its Vibe Coding product, "DaiMa," which focuses on building applications with zero barriers to entry using natural language. It targets WeChat Mini Programs and native iOS scenarios, enabling the rapid generation of ready-to-launch, real-device-usable applications. The product is geared towards individual developers and small to medium-sized businesses without professional R&D teams. Users do not need programming knowledge; they can generate complete apps by describing their needs using natural language. It also supports obtaining QR codes from real devices for testing and continuous iteration.
DeepSeek, in collaboration with Peking University, has launched DSpark, a framework for accelerating large model inference, which has been integrated into the DeepSeek-V4 series production systems. With the same total throughput, DeepSeek-V4-Flash offers a 60%–85% speedup for single-user generation, while DeepSeek-V4-Pro achieves a 57%–78% speedup. DSpark employs a semi-autoregressive architecture and a confidence-based scheduling verification mechanism, balancing draft generation speed and coherence, and dynamically adjusting the verification length based on system load.
OpenAI has released a limited preview of its GPT-5.6 series, including the flagship Sol, the balanced Terra, and the cost-effective Luna model. Sol excels in programming, cybersecurity, and biology, and adds max deep inference and ultra multi-agent collaborative modes; Terra offers equivalent performance at half the price of GPT-5.5; and Luna emphasizes lightweight and high speed. The models are now available as a limited preview due to US government requirements; they were originally planned for open access.
Mistral AI has launched its latest document recognition model, Mistral OCR 4, supporting 170 languages across 10 language families and achieving a score of 93.07 on the OmniDocBench benchmark. Compared to competitors such as GPT 5.5 Pro and Gemini 3.1 Pro Preview, the model's output is more favored by human users. The model provides bounding boxes, region classification, and confidence scores along with the output text, and supports downstream workloads such as RAG semantic segmentation.
Tencent Hunyuan open-sourced the PhoneBuddy 4B mobile agent model, simultaneously releasing five papers including PhoneWorld and PhoneHarness, covering the entire chain of environment, training, execution, privacy, and security. The model's core innovation employs a hybrid Real+Mock RL training approach, combining real apps with a self-built PhoneWorld mock environment, enabling the 4B model to outperform GPT-5.4 in multiple metrics across 150 real-device evaluations.
The DeepReinforce team has released the open-source agentic programming model Ornith-1.0, based on Gemma 4 and Qwen 3.5, and employing a self-improving training framework. The model covers four specifications from 9B to 397B, achieving state-of-the-art (SOTA) performance on the Terminal-Bench 2.1 and SWE-Bench Verified benchmarks. The flagship 397B version scores 77.5/82.4, surpassing Claude Opus 4.7; the 35B MoE version scores 64.4, surpassing Qwen 3.5-397B.
Alibaba Cloud has launched "Xiangwei," an AI-powered interactive story creation and reading platform that supports three modes: starting from scratch, continuing a manuscript, and chapter-by-chapter planning for long novels. It covers various genres including online novels, scripts, and children's stories. During reading, AI generates real-time background images, and readers can choose plot directions or customize input at key points. The platform then instantly continues the story while maintaining character consistency.
Alibaba Cloud has launched CosyVoice, a next-generation AI voice input method powered by a large-scale model. It supports real-time speech-to-text transcription, automatic filtering of redundant spoken language, self-correction during recognition, and structured text processing. The tool can understand user intent and generate standardized texts, supporting the transcription of various dialects, including Shanghainese, Cantonese, and Sichuanese, into Mandarin. The tool's main strength lies in its deep, large-scale model-driven semantic understanding and text regeneration capabilities, distinguishing it from traditional voice input methods that only perform acoustic transcription.
NetEase Youdao has released the open-source TTS model Confucius4-TTS. The model achieves three major breakthroughs: 3-second zero-sample speech cloning, accent-free synthesis across 14 languages, and emotion and prosody transfer. The model's underlying architecture employs an end-to-end structure of speech encoder + large language model + stream matching generation, with a full 54GB of weights supporting local offline deployment.
Huawei announced that HarmonyOS "Claw" is now available on all models. Users of HarmonyOS 5.0 and above only need to upgrade the Xiaoyi App to version 11.6.4.300 or higher to use it. Simultaneously, Xiaoyi Claw's service packages have been updated: the 49 RMB trial package now includes an Auto-Model mode, and the 199 RMB standard package supports selection of four models: openPangu-2.0-Pro, DeepSeek V4-Flash, DeepSeek V4-Pro, and MiniMax M3.
OpenAI has released its first self-developed inference chip, Jalapeño, in collaboration with Broadcom and Celestica. It is specifically designed for models such as ChatGPT and Codex. Led by Richard Ho, a former core member of Google's TPU team, the project was completed from design to tape-out in just 270 days, setting a new record for the fastest high-performance ASIC development. The chip improves actual utilization through optimized data handling and resource balancing, and is planned for deployment by the end of 2026, offering significantly better performance per watt than existing solutions.
Momenta, an autonomous driving company, is aiming for a Hong Kong IPO, becoming the first listed company specializing in "physical AI," officially positioning itself as a builder of foundational models for physical AI. Its R7 world model is already in mass production and deployed in vehicles, leveraging over 12 billion kilometers of real-world driving data and a three-layer architecture of pre-training, simulation, and reinforcement learning to understand and extrapolate the physical world. The company's revenue has tripled in three years, while licensing revenue has surged 42-fold, and its business model is shifting from project-based to large-scale license-based charging.
TRAE Work introduces a brand-new Design Mode, simultaneously launched on both desktop and web platforms. Users can generate design schemes through natural language dialogue, modify them in real time on the canvas, and generate interactive prototypes. This mode supports importing existing design assets such as Figma and can export design drafts to Code Mode with one click, achieving a one-stop connection from design to development.