Meituan Open-Sources 1.6-Trillion-Parameter MoE "LongCat 2.0" Under MIT License

Chinese tech giant Meituan has released the weights of its in-house large mixture-of-experts (MoE) model, **LongCat 2.0**, under the MIT license. The model is reported to have roughly 1.6 trillion (1.6T) total parameters and approximately 48B active parameters.

The weights are available for download on Hugging Face, and the local LLaMA community has already started sharing feedback. Thanks to its MoE architecture, which activates only a subset of parameters at inference time, the model is designed to maintain practical inference speeds despite its massive size.

With the open-weight model lineup growing ever deeper, it's worth noting that China's open AI ecosystem continues to expand despite U.S. regulatory pressure. In fact, a paper posted on arXiv even argues that U.S. policy has inadvertently accelerated China's open AI ecosystem.

Mistral CEO Mensch: Closed AI Models Give Labs a "Front-Row Seat to Your Business Processes"

Arthur Mensch, founder of French AI startup Mistral, has warned enterprises against dependence on closed (proprietary) AI models.

He pointed out that AI labs are accumulating ever more customer data, and that in some cases they even use that data to **compete with their own customers**. In his words, "proprietary models give AI labs a front-row seat to companies' business processes."

While the concern itself is legitimate, Mistral too struggles to compete head-on with frontier models from OpenAI and Anthropic on performance, and is effectively betting on EU sovereignty as its strategic advantage. The trade-off between data sovereignty and model performance will remain a key decision axis for enterprises considering AI adoption.

Hollywood Splits Over ByteDance's "Seedance" — Banned in Public, Used Behind the Scenes

ByteDance's AI video generation tool **Seedance** is rocking Hollywood. Viral videos of AI-generated Brad Pitt and Tom Cruise have become a hot topic, and the Motion Picture Association (MPA) has sent AI companies its **first-ever cease-and-desist letters**.

Behind the scenes, however, Simpsons animation producer Joel Kuwahara has revealed that studios are using Seedance on a "don't ask, don't tell" basis. The contradiction is now out in the open: publicly opposing AI content while quietly leveraging it to cut production costs.

It's a symbolic case of AI video generation technology starting to collide in earnest with copyright and likeness rights.

The "AI Education Divide" Seen Through Alpha School — A $75,000-a-Year AI Private School

Among affluent U.S. families, **AI-powered private schools** like Alpha School are gaining popularity. The school combines two hours of AI tutoring per day with project-based workshops, and annual tuition runs as high as $75,000.

Traditional public schools, meanwhile, are struggling to adopt AI technology, and experts warn that using AI without proper skills may actually **harm learning outcomes**. In the AI era, it is becoming increasingly clear that the quality of education is ever more determined by economic power.

"MrFlow" Speeds Up Diffusion Models 10–25x with No Training Required

A technique called "MrFlow" that accelerates flow-matching models such as FLUX.1-dev and Qwen-Image by **up to 25x** without any training is drawing attention in the community.

Using a multi-stage low-to-high-resolution pipeline, it exploits the quadratic token reduction and reduced step count of low-resolution sampling. Combined with a lightweight GAN-based super-resolution model in pixel space, it achieves end-to-end speedups of 8.25x on FLUX.1-dev and 10.3x on Qwen-Image while keeping image quality (OneIG) degradation under 1%.

What's more, combining it with timestep distillation strategies (Pi-Flow, FLUX-schnell, etc.) can deliver up to 25x acceleration. A ComfyUI plugin has also been released, making this a result with significant practical impact.

Hands-On Feedback on "Agents-A1", a Local Model for Agents

The GGUF-quantized (Q8) build of **Agents-A1**, released by InternScience, is showing practical performance on an M1 Max (64GB) machine.

Even when using the full 262K context, it records around 500 t/s for prompt processing (pp) and about 40 t/s for text generation (tg). Paired with opencode, it has been rated as running at "nearly Qwen level," making it a promising option as a local model for agents.

---