Saturday, 24 October 2026 · Europe/Bucharest · sources: Hacker News, tech press, vendor blogs, X (via Aikido posts)
A 2.8T-total / 104B-active MoE with 1M context, priced at $2/$10 per M tokens (cheaper than K3), with a 2–3x faster Highspeed tier. Open weights are promised for November; top HN story with 800+ points.
Kimi blog · MarkTechPost · HN
Wonder Full ties for second on Artificial Analysis' index (subscription only); Muse Spark Open (816B/42B active, 1M context) is now the top US open-weight model, under a license capped at $20M revenue.
Apache-2.0 MoE (23B active, 1M context) from the Nvidia-backed US lab, ranked second among US open models and notable for using fewer tokens per task; beta API via waitlist.
Built on Gemma 4 E4B with Nemotron post-training, with multi-turn tool use, configurable reasoning and 64k context, aimed at RTX, Jetson and DGX Spark devices.
Trained from scratch on human conversation audio, with no text or pretrained models, so it keeps tone and pauses that speech-to-text pipelines throw away. Waitlist only for now.
A 293B/30B-active MoE with 1M context, pitched on coding and agentic tasks and served from a 10,000-card domestic inference cluster.
It reads production logs and metrics, investigates alerts, validates a fix and, if you enable auto-publish, ships it, effectively an autonomous on-call engineer for small teams.
A workplace app-building agent that pulls requirements from mail, meetings and docs via Work IQ and can hand projects off to GitHub as draft PRs. Rolling out through the Frontier program.
A 6.3 MiB binary that writes to normal scrollback, exposes JSON output for CI, speaks the Agent Client Protocol and ships a WASM-capable library.
Google's internal design-first workflow for AI coding, with subagent personas, validation loops and review rules, released under Apache 2.0 for Antigravity.
A tmux-style tool that shows which agent (Claude Code, Codex, OpenCode and others) is working or needs you, with token tracking, per-agent change replay and worktrees.
The Brotli RCE also hits the mesh and gateway layer. Patched versions are Envoy 1.40.0, 1.39.3, 1.38.6 and 1.37.8, and Istio 1.29.9, 1.30.6 and 1.31.2, so plan ingress and sidecar upgrades.
The release candidate adds pre-routing ListenerSet filters and restores nullable retry attempts. The final v1.7 is close, with HTTPRoute retries moving to Standard.
Alpha.2 is out on Go 1.27.1. Recent merges include conditional authorization in kube-apiserver, named PodFailurePolicy rules and the removal of kubemark.
It pairs NVIDIA OpenShell and NemoClaw isolation with VAST persistent storage, so agent compute stays ephemeral while state, workspaces and audit trails survive moves. A public deploy repository is promised.
The repository for "a safe, private runtime for fleets of autonomous agents" drew about 490 points, with a lively thread on its policy and isolation model versus plain containers and microVMs.
A new post on how OpenAI isolates and hardens the environment where frontier models are trained, which matters as models get stronger at cyber tasks.
The decoder out-of-bounds write is CVSS 9.2 and works both ways, so a malicious client can hit a server and a malicious server can hit a client. Fixed versions are Node 22.23.4, 24.22.1 and 26.12.1, Brotli 1.2.1, brotlicffi 1.2.1.0 and Bun 1.4.4. Scanners don't flag the PyPI packages yet, so rebuild your images.
Aikido Security · OSV · HN
ZDI sent several reports, rated up to CVSS 9.0, to the public list instead of the security list before mainline fixes were ready. Watch for kernel updates from your distro.
The demos include a Signal 1-click and iMessage-based compromise with persistence. Details are to be released later; 800+ points on HN.
Experimental first-class type checking for scripts and a unified uv check command are a step toward one integrated Python toolchain.
Philip O'Toole's retrospective on building and running a Raft-replicated SQLite is practical reading for anyone choosing lightweight HA storage.