model

The B.AI API officially launched its first "official 10% discount group," providing users with highly reliable AI calling services through official direct connection channels while balancing model invocation stability, service quality, and cost efficiency. The first batch supports the Claude series models, with more mainstream models set to be opened successively. Meanwhile, the platform's "self-selected service provider" model offers discounts as low as 20%, granting users full autonomy to flexibly weigh performance, stability, and cost based on the actual needs of their business scenarios, achieving optimal configuration.

According to Jinshi reports, Microsoft CEO Satya Nadella emphasized that customers must have the ability to switch freely between different AI models, both closed-source and open-source. He stated during a conference call that the number of Microsoft customers using multiple models has quadrupled since the beginning of this year. Nadella pointed out that the system framework should be separate from the models, allowing for the replacement of models at any given time and the use of cutting-edge models.Microsoft has investments in both OpenAI and Anthropic, and has co-signed an open letter supporting open models with companies like NVIDIA and Palantir.

According to Jinshi reports, OpenAI announced that it will provide free access to cutting-edge models for 100,000 researchers.

According to VentureBeat, the Model Context Protocol (MCP) has received its largest update since its launch in November 2024, managed by the Agentic AI Foundation (AAIF) under the Linux Foundation. This update completes the transition to a fully stateless architecture, allowing MCP servers to be deployed behind standard load balancers without relying on "sticky routing" or shared session state, thereby simplifying large-scale cloud-native and Kubernetes deployments.In terms of security, the new version of MCP enforces validation of the issuer (iss) parameter for OAuth and OpenID Connect to prevent mix-up attacks and introduces enterprise-hosted authorization extensions to support unified identity governance for enterprises. The protocol also establishes a formal deprecation policy, ensuring that features are announced for deprecation at least 12 months before their earliest removal. MCP Apps and MCP Tasks have also been upgraded to official extensions, supporting servers in providing interactive interfaces to AI clients and executing interruptible, resumable long-running asynchronous tasks through persistent task handles. According to MCP co-creator David Soria Parra, the related SDK's weekly download volume has reached approximately 250 million, doubling in the past six months.

According to Business Insider, Amazon is making a comprehensive adjustment to its AI strategy, gradually phasing out most of its self-developed flagship Nova models, including the high-end Premier and Omni models, the Reel video generation model, and the Canvas image generation model, with the remaining products only maintaining a "Keep The Lights On" (KTLO) status. Resources are being shifted from the existing Nova series to the new frontier model project led by researcher Pieter Abbeel (who joined through the acquisition of Covariant), with the new flagship model expected to debut at this fall's re:Invent conference.This adjustment is accompanied by organizational restructuring. Amazon laid off employees in the AGI department last week and closed the AGI Lab established in 2024 by absorbing the AI startup Adept team. AGI head Rohit Prasad left in December 2025, and AGI Lab head David Luan stepped down in February of this year. The AGI organization was integrated under Senior Vice President Peter DeSantis in December of last year, merging with chip development and quantum computing. An Amazon spokesperson responded that the company "has long supported AI models in production environments" and emphasized that "AI models remain one of the company's most important work directions," with the model lineup continuously evolving based on customer needs.

Ronald Keung, the head of Asian Internet Research at Goldman Sachs, stated that as the global usage of Chinese AI models surges, developers may begin to charge commercial licensing fees to cloud platforms for hosting their open-weight models. Currently, most Chinese models are released under permissive licenses such as the MIT License, allowing any user to use, modify, and host them for free, including for commercial purposes.Keung pointed out that Chinese AI models like Kimi K3 from the Dark Side of the Moon and GLM-5.2 from Zhiyu AI have performance levels that are close to those of leading models in the United States. The adoption rate among small and medium-sized enterprises, both domestically and globally, is very fast, and even large enterprises are beginning to consider using Chinese models. Goldman Sachs believes that model developers can convert the growing use of open-source into actual revenue by requiring third-party service providers to purchase commercial licenses to run model inference on their own infrastructure.

NVIDIA released the open-source AI model NVIDIA Ising Calibration 1.5, designed for the automatic analysis of quantum processor (QPU) diagnostic data, and to autonomously determine device calibration schemes, achieving automation of the quantum computer calibration process. NVIDIA stated that Ising Calibration 1.5 is a visual language model (VLM) specifically designed for quantum computing calibration scenarios, capable of understanding experimental data from quantum chips and performing "zero-shot" analysis in the absence of historical cases, while also enabling context learning (ICL) with relevant experimental samples to help continuously optimize the operational state of quantum devices.In the QCalEval quantum calibration benchmark test, Ising Calibration 1.5 averaged about 10% ahead of similarly sized open-source models in zero-shot inference capability, and when using relevant experimental cases for context learning, it showed an approximately 86.5% performance improvement over the previous generation model, surpassing multiple open-source models and approaching the level of top closed-source large models. The model has 31 billion parameters and supports operation on NVIDIA Grace Blackwell and Vera Rubin data center GPUs. It also launched an NVFP4 quantized version, which can be deployed on a single consumer-grade GPU or NVIDIA DGX Spark, significantly lowering the usage threshold for quantum laboratories.NVIDIA claims that the training data for Ising Calibration 1.5 comes from various qubit architectures, including superconducting qubits, quantum dots, ions, neutral atoms, and helium surface electrons, providing calibration capabilities for different types of quantum computing devices. Industry experts believe that automated calibration is one of the key bottlenecks in the scalable development of quantum computing. NVIDIA's launch of this AI-driven quantum calibration tool signifies that AI models are beginning to extend from traditional computing domains into the quantum hardware control layer, potentially becoming an important component of the future quantum computing industry infrastructure.

NVIDIA CEO Jensen Huang posted a second tweet on the X platform, announcing the establishment of the "Open Secure AI Alliance," aimed at uniting global technology companies to jointly develop security technologies and tools to protect AI software and AI Agents.Huang stated that attackers already possess cutting-edge AI capabilities, and defenders need to establish a cutting-edge AI ecosystem driven by the global community, combining the advantages of both open and closed-source models to enhance AI security capabilities.He pointed out that in the previous Hugging Face security incident, closed-source AI limited critical forensic work, while open-weight cutting-edge models helped the team control the intrusion event, which is also the reason for establishing the Open Secure AI Alliance.According to NVIDIA, the alliance will collaborate with several companies and institutions, including Adobe, Microsoft, Hugging Face, IBM, NVIDIA, Cloudflare, CrowdStrike, Databricks, Palantir, Red Hat, Salesforce, SAP, ServiceNow, SpaceX, and SK Telecom, to develop new AI security technologies and protective tools.

The decentralized AI infrastructure network DGrid AI announced an official partnership with Meituan LongCat and has launched the LongCat-2.0 model under Meituan. Global developers and AI Agent builders can now access this model through the DGrid decentralized AI Gateway, providing low-cost and highly available access.LongCat-2.0 is designed for Agentic Coding in real-world scenarios, featuring a 1.6T parameter architecture with approximately 48B activations, supporting efficient long-context reasoning of 1 million tokens, and enhancing the processing speed of long text workflows through LongCat Sparse Attention. The model integrates agent, reasoning, and interaction capabilities, excelling in code understanding, generation, and execution.DGrid is a decentralized AI infrastructure that aggregates mainstream AI models globally, currently integrating over 200 models including Claude, GPT, Gemini, MiniMax, and GLM, enabling one-stop access through a unified API. The integration of LongCat-2.0 marks another expansion of DGrid on the model supply side. As more high-quality models continue to be launched, DGrid is becoming a unified entry point for AI developers to access diverse model capabilities, providing more open and flexible infrastructure support for the next generation of intelligent applications.

According to Jinshi reports, NVIDIA CEO Jensen Huang posted on social media, sharing a letter titled "Open Weights and U.S. AI Leadership," calling for the promotion of open-weight artificial intelligence model development. The letter originally published by Huang included 25 co-signers. Now, the number of co-signers has increased to 50, doubling in number.OpenAI is one of the 25 newly added organizations, along with other new organizations including Google, AMD, Cisco, Cloudflare, GitHub, Block, and Ollama. However, Anthropic and Amazon have consistently not appeared on any version of the list, and the absence of these two companies is the most noteworthy detail in the list.

NVIDIA CEO Jensen Huang posted his first personal tweet, sharing an open letter titled "Open Weights and U.S. AI Leadership," co-signed by NVIDIA and more than 20 other technology companies and organizations.The open letter argues that open weight models can expand access to the AI economy, enhance market competition, and give users more control. The signatories include Meta, Microsoft, Andreessen Horowitz, Hugging Face, IBM, Mistral AI, Palantir, Perplexity, Mozilla, Y Combinator, and others.The letter states that while open weight models carry risks, development should not be restricted through bans but should instead leverage openness to enhance AI safety and cybersecurity capabilities. Jensen Huang stated, "The world needs cutting-edge closed-source models, and it also needs cutting-edge open models."

The joint assessment by the UK Institute for AI Safety and the US Center for AI Standards and Innovation shows that Moonshot AI's Kimi K3 can assist in the development of exploit vulnerabilities and simulate cyber attacks without substantial hindrance, but its overall cybersecurity capabilities still lag significantly behind leading models in the United States.

The PPP prediction market tool monitoring shows that the probability of "the next Claude Opus model being released on July 23, 2026" on Polymarket is currently reported at 61%, up 51% in the last 24 hours. In addition, the probability of "being released before July 24" is currently reported at 9%; the probability of "being released before July 25" is currently reported at 5%.This event will settle based on the date (Eastern Time) when Anthropic's next Claude Opus model is made publicly available. Only models explicitly named "Opus" by Anthropic are valid, and they must be open to the public (including public testing or open waiting lists). Other models such as Sonnet and Haiku are not included, and the final settlement will primarily rely on official information from Anthropic.

The decentralized AI intelligent network DGrid announced that its decentralized AI model marketplace (DGrid Model Marketplace) is officially online.The marketplace is open to three types of model providers: model developers, model fine-tuners, and model deployers with computing infrastructure capabilities. They can freely list models on the platform, set their own prices, and earn real-time settlement revenue when models are called. For developers, the marketplace provides a unified entry point to discover, compare, and directly call various models through a unified API, without the need to switch between different platforms or connect to multiple interfaces.DGrid stated that the model marketplace is the "supply side" of its network, working in coordination with the AI Gateway (access side) responsible for calls, connecting AI creators and users. Currently, DGrid has aggregated over 200 mainstream models, including Claude, GPT, Gemini, MiniMax, GLM, Kimi, and has more than 15,000 paid users.In terms of quality assurance, the marketplace is supported by DGrid's self-developed Proof of Quality (PoQ) mechanism. PoQ conducts independent, random sampling of model providers through the platform's own benchmark test set and records the verification results on-chain to ensure service quality and pricing transparency—this mechanism does not touch user call data. The core members of the DGrid team have doctoral backgrounds from institutions such as Stony Brook University and have published 4 academic papers related to PoQ.Currently, the DGrid Model Marketplace is officially online. Model providers can apply to join, and developers can also experience one-stop AI model discovery and access services through the platform.

The B.AI API officially launches two models, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Among them, Gemini 3.6 Flash, as an upgraded version of 3.5 Flash, significantly improves output quality while maintaining the same pricing, with an average reduction in token consumption of about 17%. Especially in complex scenarios like DeepSWE, the maximum reduction can reach 65%, greatly enhancing development efficiency and cost-effectiveness. Gemini 3.5 Flash-Lite, on the other hand, offers ultra-fast output of up to 350 tokens/s and exceptional cost performance, precisely meeting the demands of high concurrency and high real-time tasks such as document batch processing and Agentic intelligent retrieval. Currently, both models have opened official API interfaces, allowing developers to log in to the B.AI platform immediately for a seamless experience and enjoy the performance benefits brought by the new generation of models.