Alibaba Unveils 2.4-Trillion-Parameter MoE Flagship Qwen3.8-Max, Weights Coming Next Week
Alibaba has announced its latest flagship AI model, Qwen3.8-Max. It is a 2.4-trillion-parameter Mixture-of-Experts (MoE) model, and according to reports, it is claimed to outperform Moonshot's recently released Kimi K3 on some benchmarks, putting it in the same league as frontier-class models. The company says it delivers comprehensive improvements in coding and specialized tasks, and can autonomously complete projects spanning more than 10 days.
A smaller model under the Qwen brand, Qwen3.8-27B, was announced at the same time and is also scheduled for release next week. Alibaba has indicated that it will release the Qwen3.8-Max weights for general download next week, and competition within China's AI camp — where Moonshot, DeepSeek, and ByteDance are locked in fierce rivalry — looks set to intensify further. For now, the Max version is available on Qwen Cloud.
MiniMax H3: An Omni-Modal Model That Generates Video and Stereo Audio Together
MiniMax has released H3, a general-purpose omni-modal generation system that comprehensively understands text, images, video, and audio, and generates videos up to 2K resolution and 15 seconds long with native stereo audio. The core generator, H3-Base, is a 33-billion-parameter dense transformer, notable for incorporating the pretrained weights of Qwen3-VL-32B into its encoder.
The system consists of three modules. H3-Base has been released under a community license, while H3-Context-IR, which interprets the intent of inputs, and the 2K regeneration module are currently available only via API. Stable support for conversations in 11 languages, including Arabic and Japanese, also suggests the system is built with real-world use in mind.
Devin's Maker Targets the Japanese Market: "Not Replacing Engineers, but Expanding What They Can Do"
According to Cognition AI, the company behind the autonomous AI engineer Devin, Japan has the second-largest user population after the United States. Takumi Masai, head of the company's Japanese subsidiary, said the company is paying close attention to Japan's structure as an "IT outsourcing powerhouse," where roughly 70% of IT talent is concentrated on the SIer side.
He emphasized the company's fundamental stance of "not replacing engineers, but expanding what they can do," and in the Japanese market, the company plans to focus on use cases that widen the scope of what existing IT professionals can handle, rather than simple headcount reduction. It will be interesting to see how AI agents adapt to Japan's unique structure, where outsourcing is the norm.
Running Kimi K3 Beyond Available RAM: WASTE, a Zero-Dependency C Inference Engine
From the OSS community comes an inference engine that runs the entire 2.78-trillion-parameter Kimi K3 beyond available RAM. The model's trunk stays resident in memory, only the needed experts are streamed directly from NVMe, and the remaining RAM is used as a bounded expert cache. It is implemented in C with no third-party runtime dependencies.
The approach takes advantage of the MoE-specific property that "not all experts are needed at all times," pointing to one direction for running giant models on consumer-grade hardware. It is not for everyone, but it looks like a compelling option for those who want to run some of the largest models locally.
Auditing the "Hidden Intent" of System Prompts: AISPA and SystemPromptIndex
On the research side, AISPA (Artificial Intelligence System Prompt Assurance), a new framework for auditing AI agents' system prompts from the user's perspective, has been proposed. Alongside it, SystemPromptIndex — the first and largest-scale index of AI system prompts — has also been released.
System prompts embedded in agent products are largely invisible to users and can be a source of unintended behavior. This is an effort to develop practical standards for improving transparency and accountability, and as debates over agent reliability get into full swing, it is likely to influence future governance discussions as well.