DeepSeek Open Platform Warns of Significant API Price Increase
要闻
DeepSeek posted a notice on its API interface stating that it plans to raise API service pricing across the board in the near future, with a significant increase expected. The specific pricing plan has not been announced yet; the company reminds developers to plan usage accordingly and rely on official notices.
DeepSeek has recently posted a notice on its open platform API interface, planning to increase the overall pricing of DeepSeek API services in the near future, with expectations that the increase will be substantial. The official reminder advises developers to plan their usage accordingly; the specific pricing changes and effective date have not yet been determined and will be subject to official announcements. This notice has sparked attention and discussion among community users.
Meta Releases Muse Code Beta and Muse Spark 1.2 Model
要闻
Meta released Muse Code beta and the Muse Spark 1.2 model. Muse Code is designed for large codebase tasks, capable of planning changes and coordinating multiple persistent subagents to work in parallel. Muse Spark 1.2 is an upgrade for programming built on the previous generation, improving code generation, complex debugging, codebase understanding, and end-to-end development capabilities. Muse Code is now open for installation, and Muse Spark 1.2 is available via the official API. Meta also has updates in Muse Code and the API...
Meta Superintelligence Labs has officially launched the terminal coding agent Muse Code (beta) and its driving model Muse Spark 1.2. Muse Code is specifically designed to handle complex software engineering tasks across large codebases, capable of planning changes, writing code, and verifying results, with built-in skills such as /plan, /grill, and /goal. The agent employs a joint model-agent training mechanism, coordinating multiple persistent background subagents to process complex tasks in parallel within isolated worktrees, while achieving precise recovery after crashes through local event logs. Muse Spark 1.2 offers comprehensive coding improvements and has demonstrated long-term improvement capabilities with over 1,000 tool calls in NVIDIA Hopper GPU kernel optimization tests. Muse Code can be installed with a single command on macOS and Linux, and Muse Spark 1.2 is available via the Meta Model API. Meta also offers a lower-priced Contributor tier in Muse Code and the API, with unchanged performance but collecting prompts and generated results to improve the model.
Google Restructures AI Leadership: Demis Hassabis Moves to Alphabet Chief Scientist
要闻
Google announced a restructuring of its AI leadership. Demis Hassabis stepped down as CEO of DeepMind to become Chief Scientist at Alphabet, focusing on AGI breakthroughs. Former CTO Koray Kavukcuoglu was promoted to Senior Vice President, responsible for Gemini model development.
Google and Alphabet CEO Sundar Pichai announced a major leadership change in the Google DeepMind team via the official blog. Demis Hassabis stepped down from day-to-day management duties and was formally appointed as Chairman of Google DeepMind and Chief Scientist at Alphabet, in order to focus on AGI strategy and scientific breakthroughs. He will continue to lead the AI drug discovery company Isomorphic Labs. Former DeepMind CTO Koray Kavukcuoglu has been promoted to Senior Vice President at Google DeepMind, reporting directly to Pichai, with overall responsibility for Gemini model development, frontier research, and applied teams.
Jeff Dean and Several Key Google Researchers Leave to Found AI Company Discovery Loop
要闻
Jeff Dean and several senior Google researchers announced they are leaving Google to co-found Discovery Loop, a public benefit corporation. The company aims to accelerate scientific discovery through automated machine learning and experimental loops in science and engineering.
Jeff Dean, along with Google core researchers including Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, announced their departure from Google to jointly found a public benefit corporation named Discovery Loop. According to reports, Jeff Dean will serve as CEO of the company. Discovery Loop is dedicated to using advanced AI systems to automate experiment loops, thereby significantly accelerating the speed and efficiency of research in machine learning, science, and engineering. The company has already secured support from Google parent company Alphabet as a founding investor and cloud partner, with the first round of funding co-led by Radical Ventures and Khosla Ventures.
Cloudflare Open-Sources AI 'Operating System' Cloudflare OS
开发生态
Cloudflare open-sourced Cloudflare OS, an enterprise-focused open-source AI 'operating system' platform that provides an agent workspace for everyone in an organization, based on company context and skills.
Cloudflare has released the open-source Cloudflare OS platform, providing an Agent workspace for everyone in the enterprise to build applications, automate workflows, and securely access internal systems. The platform consists of three core components: an Agent workspace grounded in company context and skills, a governance framework for secure access to internal data and services, and a personal application platform where applications can be built, shared, and continuously modified. Cloudflare OS runs on Cloudflare Workers, controls access permissions through Cloudflare Access, and manages model routing and spending via AI Gateway. Each Agent and application has no default access permissions, with resource authorization centrally managed by the Gatekeepers service. Cloudflare first deployed an internal version for employees in May of this year; the new version is now open-sourced and available on GitHub, allowing organizations to deploy it to their own Cloudflare accounts for customization.
Prime Intellect Launches Open-Source Self-Improving Agent Framework Prime Agent
开发生态
Prime Intellect launched Prime Agent, an open-source self-improvement framework. Built on recursive language models and continuous environments, it lets agents treat context as a variable to achieve self-improvement.
Prime Intellect has launched Prime Agent, an open-source agent framework designed for coding and long-running autonomous tasks. The framework is built on two core abstractions: recursive language models and a continuous environment, allowing models to treat context as a variable and make programmatic tool calls and sub-agent dispatch within a REPL. According to official data, when using the Opus 5 model, Prime Agent achieved a score of 95.5% on the ARC-AGI-3 test, surpassing the reported human expert baseline. The framework is now fully open-sourced on GitHub.
Command Code Launches GOAT Plan: $10 per Month for $70 Credit
开发生态
Command Code launched the GOAT plan, where users pay $10 per month for $70 in credit. The subscription supports over 30 open and closed models and provides global access.
Command Code has introduced a new GOAT plan subscription, aiming to provide cost-effective programming services. Users can pay $10 per month to receive $70 in credits, with the flexibility to switch between 30+ mainstream open-source and closed-source models at any time. The subscription has specific usage caps, such as a $14 limit per 5 hours and a $70 monthly limit, and available credits dynamically adjust based on the price of the selected model. The GOAT plan is currently open to global users, accessible via the Command Code CLI or Provider API, and most models offer a zero data retention option by default.
ModelScope Community Launches 'Moli' System to Replace Daily Fixed Free Quota
开发生态
ModelScope community changed the daily fixed free count for services like API-Inference to a consumable credit called 'Moli'. Users now need to log in daily to earn Moli to redeem model invocation quotas.
ModelScope community has adjusted the usage mechanism for services like API-Inference, shifting from a fixed daily free quota to consuming virtual credits called 'MoLi'. Users need to obtain MoLi by daily login, binding an Alibaba Cloud account, or participating in community contributions. MoLi is divided into short-term credits valid for 24 hours and long-term credits valid for 90 days. When consumed, the system prioritizes deducting credits closest to expiration. When using API-Inference to call models, lightweight, mainstream, and flagship models consume an average of 0.5, 1, and 2 MoLi per request, respectively. Additionally, the service requires users to bind an Alibaba Cloud account and complete real-name verification.
Meituan's LongCat-2.0 Model Available Free on OpenCode Platform
开发生态
Meituan's LongCat-2.0 model is now freely available on the OpenCode platform. However, the company has not specified the duration of the free policy or usage limits.
The LongCat team under Meituan announced that the LongCat-2.0 model is now freely available on the OpenCode platform. According to OpenCode's official introduction, LongCat-2.0 supports a 1 million token context window, adopts a fully open-source model, and is positioned as Meituan's latest model optimized for coding and code scenarios. Currently, users can use the model for free on the OpenCode platform, but the official announcement did not specify the duration of the free policy or details such as token usage limits.
Qwen Image Model Qwen-Image-3.0 Officially Launches on API
模型发布
Alibaba Qwen officially launched the API for the Qwen-Image-3.0 model, including standard and Pro versions. The official states that the Pro version ranks first in China on authoritative international benchmark leaderboards. Users can now access it via the Qwen AI platform.
Alibaba's Qwen image generation model Qwen-Image-3.0 is now officially available on the Qwen AI platform, with APIs opened to all users. The model includes a flagship version Pro and a standard version Standard, supporting 4.5k token long instruction input and native rendering in 12 languages, capable of handling complex layouts such as newspapers and exam papers. According to the official statement, it incorporates world knowledge and supports internet access for information retrieval. The flagship version Pro ranks first domestically in the latest text-to-image leaderboard on Arena.ai, with text-to-image pricing starting at 0.18 yuan per image. Overseas users can currently experience it through Qwen Cloud.
ByteDance Seed Unveils Native Audio-Visual Full-Duplex Foundation Model SeedRealtime
模型发布
ByteDance Seed officially launched SeedRealtime, a native audio-video full-duplex foundation model, now fully available on the Doubao app. The model uses a unified architecture that integrates audio, video, and text, enabling real-time interaction of 'seeing, hearing, and speaking.'
ByteDance Seed officially launched SeedRealtime, a native audio-video full-duplex large model. The model adopts a unified architecture that natively integrates audio, video, and text, enabling real-time interaction over continuous multimodal information streams, with perception, understanding, decision-making, and expression occurring simultaneously. According to official end-to-end human evaluation, compared with cascaded models, SeedRealtime reduces conversation pacing issues by half, and significantly reduces interruptions, latency, and false triggers. Currently, after updating the Doubao app to the latest version, users can experience the model in the video call interface.
Sand.ai Releases Open-Source 114B Parameter Audio-Visual Model MAGI-2 Preview
模型发布
Sand.ai officially released and open-sourced the audio-video generation model MAGI-2 Preview. The model has 114B total parameters, with only about 6B activated per token, supporting generation of 10-second 1080p audio-video clips.
Sand.ai officially released and open-sourced the unified audio-video generation model MAGI-2 Preview. The model has a total parameter count of approximately 114B, but each token only activates about 6B parameters in a single forward computation, adopting a single-stream architecture for joint modeling of text, video, and audio. In terms of generation capabilities, MAGI-2 Preview currently only supports generating 10-second video clips at up to 1080p resolution with accompanying sound. In addition, the hardware requirements for running the model are relatively high, requiring 8 NVIDIA Hopper GPUs.
JD.com Open-Sources Real-Time Video Editing System JoyAI-Video-Edit
模型发布
JD.com recently open-sourced the deployment code, online demo, and model weights for its real-time video editing system, JoyAI-Video-Edit. The system supports streaming real-time editing and achieves 30.19 FPS end-to-end at 720p resolution.
JD.com's jd-opensource released the deployment code, technical report, online demo, and model weights for JoyAI-Video-Edit, a real-time instruction-guided video editing system. The system performs causal frame processing on live streams or uploaded videos, enabling real-time editing without waiting for the complete video. According to official benchmark data, the full end-to-end pipeline can achieve a throughput of 30.19 FPS at 720x1280 resolution. The model weights are released under the Apache 2.0 license, but currently no Inference Provider offers official hosting services.
Xiaomi Open-Sources Robot Model Xiaomi-Robotics-1 Code and Checkpoints
模型发布
Xiaomi has open-sourced the code and checkpoints for its robot foundation model, Xiaomi-Robotics-1. The vision-language-action model was trained on over 100,000 hours of real trajectories and achieves state-of-the-art performance on multiple simulation benchmarks.
Xiaomi open-sourced the code and checkpoints of its robot foundation model Xiaomi-Robotics-1 (XR-1) via platforms such as GitHub and Hugging Face, including post-training, inference, and benchmark evaluation tools. The vision-language-action (VLA) model combines a pretrained Qwen3-VL with a Diffusion-Transformer, adopting a two-stage training paradigm that includes pretraining and post-training. Currently, the model is open-sourced under the Apache 2.0 license, supporting developers in custom fine-tuning and real-time deployment.
Google Labs Expands Dreambeans App to US AI Pro Subscribers
产品应用
Google Labs has made its experimental app Dreambeans available to US AI Pro subscribers, previously limited to AI Ultra users. The app uses Personal Intelligence to connect to a user's Google apps and delivers a daily personalized story collection, surfacing deep content and related topics they may have missed.
Google Labs announced the expansion of the experimental mobile app Dreambeans to AI Pro subscribers in the United States. Dreambeans uses Personal Intelligence to connect users' Google apps, providing a personalized collection of stories every day, presenting in-depth content and relevant topics that users might otherwise miss. Currently, both AI Ultra and AI Pro subscribers in the United States can use the app.
Visa Announces $2.4 Billion Acquisition of Anti-Fraud Company BioCatch
行业动态
Visa announced it has signed a definitive agreement to acquire BioCatch, a behavioral fraud prevention provider, for $2.4 billion in cash. The acquisition aims to integrate BioCatch's behavioral and device intelligence technology to help financial institution clients detect and prevent account takeover, scams, and digital fraud before payments occur.
Visa signed a definitive agreement to acquire BioCatch for $2.4 billion in cash, with funds to be paid to Permira advisory funds and other shareholders. BioCatch is a provider of behavior-based and multi-signal fraud intelligence, whose AI technology analyzes thousands of behavioral and device signals in real time to distinguish legitimate users from fraudsters. Visa expects the acquisition to complement its existing network, fraud, and security solutions, helping clients address the growing threats of account takeover and digital fraud driven by AI.
Meta Details Multi-Stage Ad Ranking Architecture, Says Conversion Rate Up 6%
技术与洞察
Meta published an article describing a multi-stage sequential model architecture for ad ranking, which decouples offline user modeling from online ranking to improve computational efficiency. The official says this technology increased Instagram conversion rates by 6% and exhibits predictable scaling laws similar to large language models.
Meta described its multi-stage sequential model architecture for ad ranking, which achieves efficient scaling by decoupling offline user modeling from online ranking. The company said the technology and related model innovations have been applied in production, driving a cumulative 6% increase in Instagram conversions, a 3% increase in Facebook conversions, and a 3.5% increase in Facebook ad click-through rates. This architecture is a core component of Meta's generative ad recommendation model (GEM), and has demonstrated predictable scaling laws similar to large language models in real traffic, with no signs of saturation thus far.
Report: ByteDance to Catch Up with Rivals Without Using Distillation
前瞻与传闻
According to reports, ByteDance's founder told an internal meeting that the company will not use distillation techniques to catch up with competitors. The report says this is to avoid triggering scrutiny, even if it means ByteDance may fall behind domestic peers in the short term.
According to The Information, the founder of ByteDance explicitly stated at an internal AI team all-hands meeting that the company will not use distillation technology to improve AI model capabilities, even if it falls behind domestic competitors in the short term. The report said the main reason for this decision is to avoid inviting a new round of scrutiny on its app TikTok. The founder told employees that the company should be willing to sacrifice short-term interests for long-term goals.
Report: DeepSeek Restarts Second Funding Round at ~¥500B Pre-Money Valuation
前瞻与传闻
Reports say DeepSeek has restarted its second funding round, planning to raise ¥50 billion. Multiple deal insiders say the round's pre-money valuation is about ¥500 billion, with signing expected by late August.
According to Caijing, multiple deal insiders revealed that the large-model company DeepSeek has restarted its second round of financing. This round plans to raise 50 billion yuan, with a pre-investment valuation of approximately 500 billion yuan, and is expected to complete signing in late August. The deal insiders said DeepSeek and investors hope the financing will be kept low-key after the restart.
Note: Content is AI-assisted and may contain hallucinations and errors.
