Xiaomi Releases MiMo-V2.5, a 311B-Parameter Multimodal Model
Xiaomi has released MiMo-V2.5, a large-scale multimodal model built on the MiMo-V2-Flash backbone, now available on HuggingFace. It is a native multimodal model that processes text, images, video, and audio within a unified architecture, and it also supports long-context reasoning and agentic workflows.
It adopts a hybrid attention mechanism, interleaving sliding-window and full attention layers, which enables efficient processing even over long contexts. The model has already racked up more than 2,600 downloads on HuggingFace, drawing attention from the community as a new option among open multimodal models.
OpenAI Lands on AWS Bedrock One Day After Microsoft Partnership Revision
Just one day after Microsoft and OpenAI dissolved their exclusive partnership, AWS began offering three OpenAI offerings on Amazon Bedrock. Among them is an agentic service jointly built by the two companies.
Until now, OpenAI's models had been available only through Azure, but this move is accelerating the trend toward OpenAI models being available across all three major clouds (Azure, AWS, and Google Cloud). Competition among cloud providers looks set to expand the range of AI platform choices.
White House Drafts Guidance to Restore Anthropic's Federal Government Access
The U.S. White House is developing guidance that would allow federal agencies to do business with Anthropic again. Access to the new model Mythos is expected to be included.
Following the earlier standoff over the Pentagon, Anthropic had been restricted from doing business with the federal government. This move appears to be part of an ongoing adjustment process aimed at balancing national security concerns with government use of AI technology. The evolving relationship between AI companies and government agencies is increasingly shaping the pace of technological development itself.
Firestorm Labs Raises $82 Million — Drone Factories in Containers for the Frontline
Defense tech startup Firestorm Labs has raised $82 million (approximately 12.3 billion yen). The company is working to realize its vision of manufacturing drones in container-sized factories and delivering them directly to the front lines.
As demand for AI-powered autonomous drones grows in the defense sector, the company is drawing attention for its approach of dramatically shortening the lead time from manufacturing to deployment. It is also an attempt to shift the traditional, centralized weapons procurement model toward a distributed one.
IK_LLAMA Adds Support for Qwen3.5's MTP (Multi-Token Prediction)
A PR adding support for Qwen3.5's Multi-Token Prediction (MTP) has been merged into ik_llama.cpp, the open-source LLM inference engine. MTP is a technique that speeds up inference by predicting multiple tokens at once, and it is expected to work synergistically with speculative decoding.
A Q8_0-quantized GGUF of Qwen3.5's 27B-parameter model has already been published, putting the local LLM community in a position to try out the practical speedups right away.
10 Million Papers Visualized in an Interactive Semantic Space Map
An interactive semantic space map has been shared with the Reddit MachineLearning community, built by embedding the 10 million most recent academic papers with SPECTER 2 and reducing the dimensions with UMAP.
It offers a fresh approach to navigating academic research, including automatic classification of semantic "neighborhoods" via Voronoi partitioning, hybrid queries combining keyword and semantic search, and ranking analyses of research institutions, authors, and topics. It uses OpenAlex as its data source and is free for anyone to use.
---