Overview
8 stories in this issue. The first 3 are today's priorities.
Hot model watch
- Top · Alibaba's Qwen open-sources multimodal plugins for six agent harnesses
- Top · Mistral releases Shieldstral, a 3B open-weight multimodal safety classifier
- Top · SaferAI: Z.ai's GLM-5.2 nears the frontier but lacks safety mitigations
- Liquid AI ships LFM2.5-2.6B for on-device agents
Global AI news 5. Amazon Bedrock adds Web Search for grounding model responses 6. Anthropic names first Chief Global Affairs Officer 7. NVIDIA joins NSF's State and Regional AI Hubs program
Regional & early signals 8. SpaceX and NVIDIA to put data-center-class compute on satellites
Jiufeng graphic based on the sources cited in this issue.
Hot model watch
Alibaba's Qwen open-sources multimodal plugins for six agent harnesses
Qwen packages multimodal capabilities as installable skills that drop into Claude Code, Codex, Gemini CLI and three more agents.
Alibaba's Qwen team has published Qwen-MM-Plugins, an Apache-2.0 open-source repository that packages multimodal capabilities as installable "skills" spanning files, media and design software. The repo documents installation for six existing agent harnesses: Claude Code, Codex, Qoder, OpenClaw, Qwen Code and Gemini CLI. The latest README update is credited to Shuai Bai, whose record includes Qwen's multimodal work such as Qwen-VL.
Per RuntimeWire, the repository tests whether multimodal tools can be installed across existing agent interfaces without adopting a separate runtime; its compatibility, external API requirements and production performance remain unverified.
Image source: GitHub; mirrored on Jiufeng R2.
Source: RuntimeWire · GitHub Qwen-MM-Plugins · Qwen team
Mistral releases Shieldstral, a 3B open-weight multimodal safety classifier
Mistral frames content moderation as policy-adaptive Q&A in a 3B model it says beats models up to 7x larger.
Mistral released Shieldstral-1.0-3B, a 3-billion-parameter, Apache-2.0 open-weight multimodal safety classifier available on Hugging Face. It frames content moderation as a policy-adaptive question-answering task: it accepts plain-language policies at inference time and unifies text and image safety evaluation without retraining, returning a calibrated safety score. Mistral is an inaugural member of the Open Secure AI Alliance formed with NVIDIA and others.
The "outperforms models up to 7x its size" claim comes from Mistral's own technical report and has not been independently reproduced; as a guardrail classifier it does not generate content, and real-world filtering depends on the configured policy.
Source: Mistral AI · Technical report · Hugging Face
SaferAI: Z.ai's GLM-5.2 nears the frontier but lacks safety mitigations
An open-weight Chinese model trails top labs by only months on cyber and bio, with a widening safety gap.
A new report from AI-safety nonprofit SaferAI finds that Z.ai's open-weight GLM-5.2 is only a few months behind OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 on cyber and bio capabilities. The report says GLM-5.2 lacks key safety mitigations, and that the divide between frontier capabilities and safety practices is growing.
This rests on a single SaferAI report as reported by TechCrunch, making it a single-reporting-source item; the write-up does not publish per-task scores, and "a few months" is a qualitative judgment.
Source: TechCrunch
Liquid AI ships LFM2.5-2.6B for on-device agents
A 2.6B on-device model, with Liquid AI's own agent-benchmark scores and local throughput figures.
Liquid AI released LFM2.5-2.6B, a 2.6-billion-parameter model built to run agents entirely on-device: it supports tool calling and multi-step workflows, and Liquid AI says it is competitive with models 4x larger on tool use, instruction following and multi-step agentic tasks. Benchmark figures published on its blog include AIME25 51.87, IFBench 59.17, BFCLv4 56.88 and ToolSandbox 77.83; on local generation throughput, Liquid AI reports about 220 tok/s on an M5 Max, 113 tok/s on Ryzen and roughly 30 tok/s on a phone.
These benchmark and throughput numbers come from Liquid AI's own blog and have not been independently reproduced; the "competitive with models 4x larger" claim does not name the compared models, and on-device performance across different hardware and quantization awaits third-party verification.
Source: Hugging Face blog
Global AI news
Amazon Bedrock adds Web Search for grounding model responses
AWS makes Web Search a generally available, server-side built-in tool on Bedrock.
AWS announced the general availability of Web Search on Amazon Bedrock, a server-side built-in tool that grounds foundation-model responses in current web knowledge. AWS says it works across chatbots, coding assistants, CLI tools and enterprise apps, helping answer questions beyond a model's training data and reducing hallucinations.
This is an AWS platform capability rather than a model advance in itself; the excerpt does not disclose pricing, latency or retrieval-quality numbers.
Source: AWS Machine Learning Blog
Anthropic names first Chief Global Affairs Officer
A former California Supreme Court justice will lead Anthropic's policy and government relations worldwide.
Anthropic said Mariano-Florentino (Tino) Cuéllar will join as its first Chief Global Affairs Officer, leading policy, strategic international engagement and government relationships worldwide. Cuéllar, a former California Supreme Court justice, is being brought into the leadership team by Anthropic co-founder and president Daniela Amodei. RuntimeWire frames the move as coming while Claude reaches more large companies and public institutions.
This is a personnel and policy change rather than a model update; the exact scope and agenda are per Anthropic's announcement.
Source: Anthropic · RuntimeWire
NVIDIA joins NSF's State and Regional AI Hubs program
NVIDIA is participating in a newly launched US program to expand AI computing for research and education.
NVIDIA said it is participating in the U.S. National Science Foundation's State and Regional AI Infrastructure Hubs program, launching now to expand access to advanced computing, data, software and expertise for AI-enabled research and education. NVIDIA says the program is consistent with the aims of the Genesis Mission and will support multistate groups of colleges and universities.
This is a US national program aimed at research and education; the excerpt does not disclose funding size, the full list of participating institutions or a timeline.
Source: NVIDIA AI Blog
Regional & early signals
SpaceX and NVIDIA to put data-center-class compute on satellites
SpaceX and NVIDIA are co-developing the Starmind AI1 payload, with each satellite carrying an NVIDIA Rubin GPU and Vera CPU.
Chinese-language reporting (Yicai, 36Kr) says that on August 5 Beijing time, SpaceX announced on social media that it is working with NVIDIA to co-develop the Starmind AI1 satellite compute payload; each Starmind satellite would carry an NVIDIA Rubin GPU and Vera CPU to deliver "data-center-class" compute in space. NVIDIA confirmed the collaboration on social media, calling it "the next chapter of AI infrastructure."
Evidence boundary: this is currently carried only by Chinese-language outlets (Yicai, 36Kr) relaying the two companies' social-media statements; payload specs, deployment timing and compute scale are not yet confirmed by official technical documentation or independent English reporting.
Get the latest AI model insights and tutorials from Jiufeng.
Explore more
