MiniMax Releases M3, a 427B-Parameter Natively Multimodal Model
Chinese AI company MiniMax has released "**MiniMax-M3**", a natively multimodal model boasting approximately 427 billion parameters (around 23 billion active), on Hugging Face. Its standout feature is the ability to process text, images, and video in an integrated way. Because it was trained on mixed modalities from the very first stage, it is said to achieve deeper semantic fusion than conventional "bolted-on multimodal" approaches.
The biggest highlight is its **1 million token context length**. Thanks to a newly developed sparse attention mechanism called "MiniMax Sparse Attention (MSA)", the per-token computational cost at 1 million context has been reduced to roughly 1/28.4 of conventional GQA. In real-world measurements on H800 GPUs, the model achieves a 14.2x speedup in prefill processing and a 7.6x speedup in decoding. MSA is a block-level sparse attention mechanism built on Grouped Query Attention: a lightweight Index Branch scores KV blocks and selects a Top-k subset for each GQA group.
M3 ships with two inference modes: "thinking" and "non-thinking". The thinking mode is aimed at complex reasoning and agentic tasks, while the non-thinking mode targets low-latency scenarios such as chat and code completion. A technical report (arXiv:2606.13392) has also been published, and the model is already available on HuggingChat.
Notably, the paper also mentions a 109B-parameter model, fueling expectations for the release of a smaller version capable of running on consumer GPUs.
Google and FBI File Their First Joint Lawsuit, Targeting a Chinese AI Fraud Network
Google has filed its first joint civil lawsuit together with the FBI, pursuing AI abuse by the China-based cybercrime network known as "**Outsider Enterprise**". The network allegedly used AI tools such as Gemini to generate and send approximately 2.5 million fraudulent messages.
At nearly the same time, OpenAI announced that it had identified and blocked two Chinese influence operation clusters, both of which reportedly targeted US infrastructure and political discourse. The fact that Google and OpenAI independently exposed China-origin AI abuse at almost the same moment is telling. The risk of AI being exploited for nation-scale information warfare is becoming more real than ever.
This Google-FBI joint lawsuit could set a precedent for private tech companies and law enforcement agencies joining forces against AI-driven crime, and it may well influence future cybersecurity policy.
Bernie Sanders Proposes an "AI Sovereign Wealth Fund"
US Senator Bernie Sanders has unveiled a proposal for an **AI Sovereign Wealth Fund** designed to return the benefits of AI to all citizens. Prominent security researcher Bruce Schneier has analyzed the proposal in detail on his blog, and it has also sparked discussion on Hacker News.
The core idea is to share the productivity gains and economic benefits that AI delivers with the nation as a whole, rather than with private companies alone. The plan envisions applying a mechanism similar to Norway's oil fund to AI, accumulating taxes and profit distributions from AI-related companies into the fund and then distributing them to citizens.
With the rapid advance of AI fueling concerns about employment, how to design frameworks for worker protection and benefit distribution has become a major policy issue in many countries. Whether this proposal will ever turn into actual legislation remains uncertain, but the fact that the debate over wealth redistribution in the AI age has entered mainstream US politics is noteworthy.
Microsoft Advocates "Spec-Driven Development" as a New Approach to AI-Native Development
Microsoft has introduced the concept of "**Spec-Driven Development**" on its Dev blog. It is a development methodology for AI-native software engineering in which you first define a "Spec" (specification) and then have AI generate code based on that specification.
Microsoft points out the limitations of the conventional approach of writing prompts and letting AI generate code, and proposes placing structured specification definitions (Specs) at the center of the development process. Microsoft also provides a toolkit called "GitHub Spec Kit", which enables systematic writing, managing, and communicating of specifications to AI.
This suggests a transition from the stage of "having AI write code" to the stage of "collaboratively designing specifications with AI and driving development from those specs". It is a symbolic move, showing that software development best practices are being fundamentally reexamined for the AI era.
Nvidia Partners with Abridge to Co-Develop AI Healthcare Models
The Wall Street Journal reports that Nvidia has teamed up with **Abridge**, a healthcare AI startup, to co-develop AI models specialized for the medical field.
Abridge builds AI systems that record and summarize physician-patient conversations in real time. The partnership aims to leverage Nvidia's GPU infrastructure and AI technology foundation to accelerate practical AI adoption in clinical settings. Healthcare is one of the fields where AI applications hold the greatest promise, while at the same time imposing extremely strict privacy and safety requirements. Nvidia's deepening investment in healthcare reflects the broader trend of industrial AI applications expanding from consumer-facing products into specialized domains.
---
**Summary:** This round-up spans a remarkably wide range of topics: the arrival of an open, giant multimodal model (MiniMax-M3), AI as a national security risk (the Google/FBI lawsuit), a vision for returning AI's gains to society (the Sanders sovereign wealth fund), the evolution of development methodology (Spec-Driven Development), and the deepening of industry-specific AI applications (Nvidia × Abridge). It shows that alongside the progress of AI technology itself, its impact on social institutions and development culture is unfolding in parallel.