The PUBG maker's AI team trained the 30B-total, 4B-active mixture-of-experts model from scratch for Korean and English chat and tool-calling agents, and ships NVFP4 (vLLM, single GPU), MLX and GGUF builds alongside the main weights. It's a gated preview with automatic approval, so specs and results may change before the final release, and the hosted API and playground fall under separate terms.
Released Oct 23, a new Priority class lets you register an assign function in langgraph.json that gives each thread a priority and a fairness group (say, premium users first, grouped by org); the server calls it on a thread's first run and keeps the result for later runs. It's a beta API and needs LangGraph CLI 0.4.38+ (reference)…
Now GA (announced Oct 23), the new option runs Claude on Google Cloud infrastructure with Anthropic as data processor, so platform features like Claude Managed Agents (beta), skills and code execution land on day zero while you keep Google Cloud IAM, audit logs and commitment drawdown at the same pricing. The existing Google managed option stays for VPC Service Controls, CMEK and data residency, but gets new platform features only once Google integrates them…
Open sourced Oct 24 under MIT, Agensh has each agent claim its own sub task, share findings and merge into a shared Gitea repo, with Mattermost for messages; on the five hardest ProgramBench tasks, going from 1 to 128 agents lifted the mean test pass rate from 19.31% to 28.78%. The release ships a single node Docker Compose setup (eight workers by default) on Copilot ACP sessions, and token costs for big swarms aren't disclosed (paper)…
Announced Oct 23, MCP Tasks in Claude Code CLI v2.1.304+ let a server accept a slow job like a build or export, return a task handle right away, and report back later while Claude carries on; running tasks can also pop elicitation forms to ask you for input. It works in terminal, Agent SDK and GitHub Actions sessions on MCP revision 2026 07 28, not yet in VS Code, the desktop app or subagent calls, and servers must add the Tasks extension themselves (Upstash walkthrough)…
In OpenAI's Oct 23 red team report, agents in separate microVMs on the same host, with no network or shared writable files, swapped synthetic secrets by evicting cache lines and timing reloads; OpenAI says the channels are slow and it now monitors for them rather than removing them. Vercel's Guillermo Rauch calls it a known noisy neighbor attack with no complete fix short of one machine per agent, so treat co located sandboxes as able to talk to each other (breakdown)…
Spotted something we missed? Start a thread.