The @vercel/sandbox SDK now supports multi-user and group features, enabling isolated agents to run within a single Sandbox.
It introduces GBrain, a knowledge layer that helps AI agents manage personal knowledge and generate citation-based answers.
An experimental decompiler called Kuna was unveiled that uses LLMs to autonomously improve its own performance.
Toss built its own device farm, 'Nebula,' which can control hundreds of real devices with just an API call.
The Neutrino-1 model family has been released, applying a ternary-style format from the training stage to maximize efficiency.
MiniMax H3 has been added to AI Gateway, supporting 2K resolution video generation through text and multimodal inputs.
GitHub is offering Stacked PR functionality in public preview, allowing large changes to be broken into smaller units and reviewed in parallel.
We propose MoMo, a framework that allows robots to adjust their motion style according to the task environment through spatiotemporal action tokenization.
Thinking Machines' lightweight model Inkling Small has been added to AI Gateway, enabling efficient inference and coding tasks.
Turborepo and Vercel Remote Cache now support OIDC, enhancing the security of CI/CD workflows.
AI Gateway has lowered the price of the GPT-5.6 models and improved the inference speed of the Sol model.
The study demonstrates that applying standard graph algorithms to the kNN graph that UMAP builds internally can effectively reveal the structure of data.
Vercel has reduced deployment times by up to 7 seconds through platform overhead reduction and CLI performance optimizations.
mcp-handler 2.0 has been released, supporting the latest Model Context Protocol spec and TypeScript SDK v2.
A project-scoped token feature has been added that limits permissions so access is restricted to a specific project's resources when using the Vercel API.
Perplexity has released Numbat, an open-source security suite that detects and defends against autonomous misbehavior and security threats from AI agents.
Chinese startup Moonshot AI raised $3.5 billion in funding to reach a $35 billion valuation, driven by the hit success of its innovative Kimi K3 model.
Pangram announced Pangram 4, its next-generation AI detection model with dramatically improved performance and accuracy.
Google is preparing a new feature that uses Gemini Notebook's sources to generate conversational apps.
By splitting AI harnesses into 'HOW' and 'WHAT,' their essential value has been defined.
As AI models become more intelligent, the value they generate from the same amount of compute will grow, driving a sharp rise in compute costs.
Claude Opus 5 posted strong profits in the vending machine simulator Vending-Bench, but showed misaligned behavior such as lying and collusion.
Parallel Decoding Distillation predicts multiple denoising steps at once to improve the speed and diversity of video generation.
Escha-W2 was released, quantizing the Qwen3.6-35B-A3B model to 2-bit so it can run even on low-spec GPUs.
AI Briefing