Today's highlight ·
Anthropic releases sandbox-runtime for OS-level agent sandboxing
Anthropic has released sandbox-runtime, an OS-level sandbox for AI agents that works without containers, MarkTechPost reports.
News
One AI innovation a day · open models, agents, tools for teams · with sources · by Kalpak Kale
News is a daily update on one real AI innovation, drawn from primary sources and linked so you can read them yourself. The newest update is the day's highlight until the next one lands. 10 updates, newest first. Longer, hand-written pieces are on the Blog.
Moonshot AI Kimi K3.1 Pro nears frontier, open weights due Nov 4
Moonshot AI's Kimi K3.1 Pro scores 65.7 percent on Terminal-Bench 4.0, closes in on Claude Opus 5.5, and promises open weights by November 4.
NVIDIA opens OpenShell, a safe runtime for autonomous agent fleets
NVIDIA's OpenShell gives agents file, package, API and credential access under kernel enforced policy and formal verification.
Recursive launches Flint, an open local coding model for agents
Recursive Flint: a free 40B MoE coding model with 2.5B active parameters, 74.2% on SWE-Bench Verified, 125 tokens a second on an M4 Pro MacBook Pro.
Google and Meta open-source TorchTPU for native PyTorch on TPUs
TorchTPU explained: Google and Meta's native PyTorch backend for TPUs with eager mode and torch.compile, replacing PyTorch/XLA. What changes for ML teams.
MiniMax M3.1 Flash undercuts DeepSeek on cost and coding tasks
MiniMax M3.1 Flash: a 428B MoE agentic model with 1M context and open weights, scoring near DeepSeek V4.1 Flash on coding at about half the cost per task.
Sphinx quizzes you on AI-written code before you push
Sphinx is a free open-source Git hook that quizzes you on your AI-generated code before push. It works with common coding agents and logs skips.
RealtimeBench introduces 180 tasks to test agents on a live clock
RealtimeBench adds 180 live-clock tasks and shows faster agents can beat stronger ones, with TORSI lifting a fast student from 13.51 to 24.19.
NVIDIA adds Nemotron 3.5 Super VL and Bolt for agent workloads
Nemotron 3.5 Super VL explained: NVIDIA's 120B open multimodal agent model, plus Bolt 74B and Bolt E4B for local and edge agents, and what each is for.
Meta and Apple ship Muse Spark Open for desktop Macs
Muse Spark Open on Mac: Meta and Apple's 816B open multimodal coding model with 1M context, running locally on two M5 Ultra Mac Studios. Features and limits.
