Mistral Releases "Medium 3.5" with 128B Parameters
Mistral has released its latest model, "Mistral Medium 3.5." The model, featuring 128B parameters, has been published on HuggingFace and is drawing significant interest from the community. Covered by multiple sources, it appears positioned to strengthen Mistral's existing model lineup.
Mistral has earned strong recognition for its open-weight models, and with the release of Medium 3.5, the options expand beyond API access to local inference as well. Detailed benchmark results await further analysis, but the 128B scale is presumably aimed at striking a balance between practicality and performance.
IBM Releases the Granite 4.1 Family (3B/8B/30B)
IBM has released "Granite 4.1," an open-source model family. Available in three sizes—3B, 8B, and 30B—it covers a wide range of use cases from edge devices to servers.
The models are available on HuggingFace and positioned as part of IBM's enterprise AI strategy. The Granite series is offered under a commercially usable license, and adoption in enterprise environments is anticipated.
Cognition in Talks to Raise Funds at a $25 Billion Valuation
It has been reported that Cognition, an AI coding company, is in talks to raise funding at a $25 billion (approximately 3.8 trillion yen) valuation. Backed by growth expectations in AI-powered code generation and development automation, the company's valuation has been climbing rapidly.
While numerous players such as Cursor and GitHub Copilot compete fiercely in the AI coding space, Cognition's autonomous AI engineer approach, exemplified by Devin, has captured investor attention. That said, some observers believe its practicality and profitability in real-world development settings remain unproven.
Claude AI Agent Deletes Company Database — Risks of Autonomous AI Operations Come to Light
A case has been reported in which a user, leveraging Claude AI's agent capabilities, accidentally deleted a company database in just 9 seconds. PocketOS has also reported losing all of its data due to an AI-related incident.
As AI agents grow increasingly capable of autonomously performing system operations, the risks of running them without proper safety measures (guardrails) have become a reality. As flagged in a previous report (76% of tool calls have no guardrails), establishing governance frameworks for AI agents has become an urgent priority.
Google Expands AI Features to Photos and Gmail
Google Photos has announced a feature that uses AI to recreate the iconic closet from the movie "Clueless." An experiential feature combining image recognition and AI styling, it demonstrates the expanding breadth of consumer-facing AI applications.
It has also been reported that AI Overviews will be extended to work accounts in Gmail. The feature, in which AI summarizes and analyzes email content, is expected to further advance the use of AI as a productivity tool.
Anaconda Acquires Outerbounds to Consolidate Its AI Development Platform
Anaconda, known for its Python distribution, has acquired Outerbounds, an ML infrastructure company. The move aims to comprehensively cover the AI development lifecycle, from building data science environments to managing machine learning pipelines.
As AI development toolchains grow more fragmented, demand is rising for integrated platforms that deliver an end-to-end development experience. Attention is focused on whether the combination of Anaconda's existing user base and Outerbounds' technical expertise will become a key differentiator against competitors.
HuggingFace Analysis: AI Evals Emerge as a New Compute Bottleneck
An analysis published on the HuggingFace Blog argues that evaluating AI models (evals) is becoming a new bottleneck for compute resources. As model performance improves, the compute required for comprehensive evaluation has surged, with some cases reportedly requiring as much compute as training itself.
The analysis highlights the reality that evaluation infrastructure has failed to keep pace with the accelerating speed of model development. Making evaluation more efficient and automated has become a critical challenge in optimizing the entire AI development pipeline.
---