Overview
8 stories in this issue. The first 3 are today's priorities.
Hot model developments
- Top · Ant unveils Ling-3.0-flash, a 124B hybrid-reasoning model
- Top · Claude Opus 5 launches, nearing Fable 5 at half the price
- Top · GPT-5.6 Sol, Terra and Luna go GA on Amazon Bedrock
Global AI news 4. Nvidia, Microsoft and Meta oppose broad open-weight restrictions 5. Google publishes a GKE AI security blueprint with open-source AIBOM 6. Meta turns its AI chatbot into an assistant with calendar briefings
Regional and early signals 7. China's UniWorld-View world model tops a Fei-Fei Li team benchmark 8. Xiaomi phone filing bundles MiMo, DeepSeek, Qwen and more
Jiufeng graphic based on the sources cited in this issue.
Hot model developments
Ant unveils Ling-3.0-flash, a 124B hybrid-reasoning model
Ant's Bailing team releases Ling-3.0-flash, using 124B total / 5.1B active params to match its previous 1T flagship.
On July 24, Ant Group's Bailing (Ling) large-model team released Ling-3.0-flash, a native hybrid-reasoning model. Ant says it carries 124B total parameters (about 12.4% of the previous 1T-class flagship reasoning model Ring-2.6-1T) and only 5.1B active parameters (about 8.1% of Ring-2.6-1T), yet matches or exceeds Ring-2.6-1T across the board. For reference, Ring-2.6-1T is open-weighted on Hugging Face and ModelScope and ships an adjustable Reasoning Effort control with "high" and "xhigh" levels.
Limitations: the announcement gives parameter counts and a "matches the prior flagship" claim but no named benchmarks or scores, and does not state whether Ling-3.0-flash itself is open-weighted or only served on Ant's platform; the comparison is the vendor's own and awaits third-party evaluation.
Image source: huggingface; mirrored on Jiufeng R2.
Source: IT Home (Chinese-language source) · Hugging Face (Ring-2.6-1T)
Claude Opus 5 launches, nearing Fable 5 at half the price
Anthropic ships its fifth-generation flagship Opus 5, which it says approaches Fable 5's frontier intelligence at roughly half the price.
Anthropic released Claude Opus 5 on July 24, the first model in the fifth Opus generation and billed as its most capable Opus. Anthropic says Opus 5 comes close to Claude Fable 5's frontier intelligence at about half the price, sets a new state of the art on coding and knowledge-work evaluations such as Frontier-Bench and GDPval-AA, and becomes the new default model tuned for everyday, efficient use across agentic coding, knowledge work, visual understanding and long-running tasks. It is available today on Amazon Bedrock and on Claude Platform on AWS.
Limitations: Anthropic states Opus 5 still trails Mythos 5 on cybersecurity tasks; the "half price," "near Fable 5" and "state of the art" claims are the vendor's own evaluation results, with no score margins or independent replication yet.
Source: Anthropic · AWS Machine Learning Blog
GPT-5.6 Sol, Terra and Luna go GA on Amazon Bedrock
OpenAI's GPT-5.6 variants Sol, Terra and Luna are now generally available on Amazon Bedrock.
AWS announced on July 24 that three OpenAI GPT-5.6 variants — Sol, Terra and Luna — are generally available on Amazon Bedrock. They are accessed through the Responses API on the bedrock-mantle endpoint, and AWS published guidance on selecting a model and reducing cost.
Limitations: this is an AWS hosting-availability and how-to post; it does not disclose each variant's capability positioning, context length or pricing. It is a model-availability milestone, so actual performance should be judged from the official model cards and hands-on tests.
Source: AWS Machine Learning Blog
Global AI news
Nvidia, Microsoft and Meta oppose broad open-weight restrictions
A group of AI firms signs an open letter against "premature" broad limits on open-weight models; Jensen Huang backs it with his first X post.
Hugging Face, Meta, Microsoft, Mistral and Nvidia signed an open letter urging policymakers not to impose broad "premature restrictions" on open-weight AI models. It arrives as Washington debates how to respond to allegations that Chinese AI labs are stealing US intellectual property; the letter never mentions China, framing open weights as key to US competitiveness, economic opportunity and national security. Nvidia CEO Jensen Huang used his first-ever X post to sign and endorse it, joining leaders such as Satya Nadella.
Limitations: this is an industry lobbying letter, not a regulatory instrument; Anthropic is not among the signatories. Mistral is one of the few non-US signers, and the framing centers on "American technological leadership."
Source: TechCrunch · Open letter PDF (Nvidia) · Jensen Huang on X
Google publishes a GKE AI security blueprint with open-source AIBOM
Google Cloud lays out a three-layer security plan for AI workloads on GKE and open-sources a controller that auto-generates an AI bill of materials.
Google Cloud released a security blueprint for CISOs and platform teams, proposing a three-layer plan covering infrastructure security, model integrity and application security, arguing that AI apps move from prototype to production faster than traditional security models can keep up. At the infrastructure layer it recommends Confidential GKE Nodes, extending hardware memory encryption to Nvidia H100 GPUs and TPUs; at the model layer it introduces k8s-aibom, an open-source Kubernetes controller that auto-generates an AI bill of materials (datasets, ML frameworks) that a traditional SBOM misses; at the application layer, Model Armor detects threats such as prompt injection.
Limitations: the blueprint is Google's recommended practice-plus-product bundle, and deployment still depends on the GKE stack; it was authored by Glen Messenger and Shannon Kularathna, and results must be validated per environment.
Source: InfoQ (Chinese-language source) · k8s-aibom (GitHub)
Meta turns its AI chatbot into an assistant with calendar briefings
Meta adds productivity features to Meta AI, letting it read your calendar to generate daily briefings, positioning it against Gemini, ChatGPT and Claude.
The Verge reported on July 24 that Meta is upgrading Meta AI from answering questions and generating images into something more like an assistant: the latest update lets it tap your calendar to help plan and generate daily briefings, framed as a productivity tool competing with Gemini, ChatGPT and Claude.
Limitations: the report focuses on the feature direction and does not specify rollout regions, the calendars/apps it connects to, or availability timing; these are product-level capabilities with no independent testing yet.
Source: The Verge
Regional and early signals
China's UniWorld-View world model tops a Fei-Fei Li team benchmark
Rabbit-pre (Tuzhan) with Peking University and Pengcheng Lab tops a novel-view-synthesis leaderboard with UniWorld-View; code and weights are fully open and Ascend-ready.
Per QbitAI, a world model called UniWorld-View, built by Tuzhan together with Peking University and Pengcheng Laboratory, topped a leaderboard associated with Fei-Fei Li's team. From a single monocular video or even one image, it generates novel-view video with a specifiable camera trajectory (pan, tilt, dolly, or a full 360° orbit), emphasizing precise camera-pose control; training data comes from PKU-linked startup Yuankong, and code and weights are fully open-sourced at PKU-YuanGroup/UniWorld-View, already adapted to domestic Ascend accelerators.
Limitations: this is a Chinese-language source; the report does not quantify the leaderboard name, scores or baselines, so the "tops the leaderboard" claim awaits the original benchmark and third-party reproduction.
Source: QbitAI (Chinese-language source)
Xiaomi phone filing bundles MiMo, DeepSeek, Qwen and more
A leak says a new Xiaomi phone filed roughly ten large models on a single device, spanning its own MiMo and several third-party models.
Citing leaker @数码闲聊站, IT Home reports that a new Xiaomi phone (model 2608BPX34C) filed about ten large models at once, including Xiaomi's own model, Xiaomi HyperVision/HyperPerception models and MiMo, plus third-party models such as Zhipu AI, ERNIE Bot, Yunque (Doubao), DeepSeek, Qwen and Hunyuan — likely tied to a new HyperOS 4 device.
Limitations: this is a Chinese-language source relaying a leaker's claim, not an official Xiaomi announcement; the exact device, launch timing and how each model is integrated remain unconfirmed.
Get the latest AI model insights and tutorials from Jiufeng.
Explore more
