Google opened Gemini 4 Argon to every paying Gemini API customer on Oct 13 (announced Oct 14), aimed squarely at long-horizon agentic work: code migration, multi-step research, cybersecurity. The headline is the price — $2 in / $10 out per million tokens, 95% off cached input — which is Opus-class money at a fifth of the cost. The catch, per Reuters/VentureBeat: it doubles to $4/$20 on Dec 1, matching Claude Opus 5.5 exactly. Seven weeks of arbitrage, then parity.
Specs that matter: 1M combined context with up to 1M output tokens (per MarkTechPost’s breakdown), sparse MoE, model ID gemini-4-argon. Already live in Cursor, Devin, and GitHub Copilot; Google claims Antigravity pilots raised agentic code submissions 35% while cutting rollbacks by a third. For the Hermes stack: nothing here is local, but Argon is now the obvious heavyweight to benchmark a hard task against when the local lane stalls — just don’t architect anything on $2/$10.
Open-sourced Oct 14: a TypeScript framework where the human UI, the HTTP API, the CLI, MCP, and A2A all share one set of actions, data, and application state. No separate “agent mode” bolted on afterward — the agent and the person literally drive the same controls. That’s the architecture the Hermes desktop/phone console has been groping toward from the other direction, and it’s a cleaner mental model than exposing everything as MCP tools and hoping for the best.
Caveats apply: beta, production reliability unproven, and the hosted-workflow upsell is the business model. But the repo is the thing to steal ideas from, not necessarily the framework to adopt. Sources: RuntimeWire (via CEO Steve Sewell’s X thread).
Released Oct 8 with Apache 2.0 weights: trained with repository-level reinforcement learning to explore code, edit files, and check its own changes, with a small active parameter count. Translation: purpose-built worker material — the classification/compression/coding-subagent grunt work in a multi-agent harness. Apache 2.0 means no license gymnastics to self-host it on the box. Watch the memory requirements before assuming “small” means “fits anywhere” — the throughput numbers were measured on server hardware.
This is the one. Releasing today, Oct 15: a 74B-parameter mixture-of-experts with only 9B active per token, compressed via quantization-aware training to a 24GB footprint, purpose-built for local agents (coding, research, tool use). Fully open source — weights, training data, recipes — and wired into llama.cpp, vLLM, Ollama, Unsloth, and LM Studio at launch, plus Hermes and OpenClaw integrations. A smaller Bolt E4B follows Oct 23 for conversational agents and game NPCs (demo: SkyrimNET). The MoE trick is the whole story: decouple total knowledge from per-token compute, and suddenly “frontier-class” fits on consumer VRAM.
The sleeper in the same announcement is PAIR going GA (Oct 13): NVIDIA’s free, open-source router that auto-discovers GPUs on your LAN and distributes inference across them, now with llama.cpp support and direct Hugging Face downloads, on Windows/Linux/macOS. For this household that’s the mini PC + the ghost box + the main rig acting as one cluster. Also shipped: Llama for Windows (free OSS llama.cpp GUI), Broadcast 2.3 with a built-in MCP server, and Switchyard, an open-source router that shifts queries between local and cloud models to minimize the cloud bill. RTX Spark laptops (Grace CPU + Blackwell GPU, up to 128GB unified) go on sale tomorrow, Oct 16.
Honest caveats: the speed-vs-intelligence claims are NVIDIA’s own — independent benchmarks arrive once the weights land. And 2-bit quantization is aggressive; quantization-aware training helps, but trust-but-verify on your own trajectories. The calculator below answers the real question: what does a given model actually cost you in VRAM?
Announced Oct 12: a desktop built on the Ryzen AI Max+ PRO 495 (16 Zen 5 cores, 55W) with up to 192GB of LPDDR5x-8533 unified memory at 273GB/s, up to 160GB of which can be allocated to the iGPU as VRAM. AMD’s vendor claim: 300B+ models at 4-bit. That number needs independent speed verification — bandwidth is the ceiling and 273GB/s is merely decent — but the form factor matters: a quiet box that can hold your entire local stack in memory at once, no spillover games. The 16GB Arc A770s keep their day job; this is the “one box that runs everything” alternative.
Source: e-bitco.in (via AMD’s Oct 12 blog post by Warren Eng).
Named at the OCP summit Oct 13: a PCIe inference card on Xe3P with up to 1.3 PFLOPS FP8, 1.5TB/s bandwidth, and — the actual headline — up to 480GB of LPDDR5X in a 350W air-cooled card. Intel isn’t chasing training; it’s wagering that production LLM serving is a memory-capacity and deployment-economics problem. Datacenter-only and nowhere near a home lab today, but file it under “2028’s used-market dream card” — the inference-economics logic is exactly right for agents that live on long contexts.
Released Oct 14, timed to KDE’s 30th anniversary: the first Plasma release with no X11 desktop session at all — Wayland is the desktop, X11 apps ride XWayland. Also in the box: the second half of the Union theming tech preview (now styling classic QtWidgets apps like Dolphin and Kate, not just QtQuick), the Kup backup utility folded into the project proper, a revived Oxygen theme, and a KWin commit_timing protocol for real frame-latency measurement. KDE even shipped a feature-length documentary for the anniversary.
For an Omarchy/Hyprland household this is spectator news, but the direction matters: the Linux desktop just voted, decisively, that X11 is a museum piece. And Union’s “one theme, both toolkits” goal is the idea every ricing setup eventually reinvents. Sources: Linux Compatible, The Register.
Built Oct 12, signed by Jan Alexander Steffens, and sitting in core-testing as of Oct 13: the next kernel headed for your pacman -Syu. Nothing to do yet — testing repo, not stable — but the usual drill applies: it’ll ride the conveyor into core within days, and Arch users on the 9800X3D box get it with the next full sync. Boring is good; boring is how kernels should arrive.
Out in the last couple of weeks: C variadic functions are now stable Rust, alongside new stable APIs. The practical upshot is cleaner FFI against C libraries that take ... — no more VaList contortions in unsafe blocks when you’re writing the little native shims that agent tooling keeps needing. Incremental, yes; but the kind of incremental that removes a papercut you hit monthly.
After nearly three years in early access (Jan 2024 → today), Keen Games flips the switch at 08:00 PDT: PC via Steam and PS5, $39.99 with a 10% launch discount. The 1.0 update completes the Embervale map (Cinder’s Cradle floating-island biome), closes the storyline, raises the cap to 50, and adds Sparkworks — a logic-based creation toolkit with Adventure Sharing. 5.5M+ players since EA. The rare survival game that took the time instead of shipping a roadmap. PS5 Pro gets a 60fps PSSR 2.0 mode; crossplay is not in at launch; Xbox waits until spring 2027.
Today’s v0.11.0 mini-update adds Lua script mod support to the ModKit — and, much more interestingly, an in-game MCP that lets you build those mods with an LLM, plus new CLI commands (max cars, max Zois) you can invoke in plain English: “make the maximum number of Zois in the city 10.” A life sim with a native agent-control plane is the “AI plays the game” demo worth watching — NPC automation, cheat scripting, and mod generation all through the same protocol Hermes speaks. Someone’s going to build a Zoi that runs errands; the only question is who first.
Season 10.5’s Halloween event is live: three weeks of event missions and a Pop & Win balloon board, with the free event-pass track topping out at a Rocket Raccoon costume. Paid track carries Scarlet Witch and Hela legendaries. Any non-competitive mode counts toward progress, so it’s a casual grind, not a ranked derailer. Log in, play quick match, collect raccoon.
RouteWeaver: cache-safe speculative inference + CPU/GPU transfer tuning for a 27B Qwen on a 12GB RTX 3060 — 48.6 tok/s in the Oct 8 experiment. This is your talker-lane model on weaker VRAM than yours; the techniques transfer straight to the dual-A770 setup.
The cross-platform “agent that actually executes tasks on any OS” — top of the Oct 11–12 star-velocity lists. Worth a look for execution-sandbox ideas and for how it handles the Hermes-style always-on loop.
The framework behind Beat 01’s story: apps where humans and agents share one action layer across UI, HTTP, CLI, MCP, and A2A. Steal the architecture diagrams.
AI agents that analyze native binaries, RE support in multiple languages. Small but directly on the REA money thread — a working reference for the agent-assisted RE workflow you’re scoping.
Turns a decade of researcher experience into usable AI “skills” guiding idea-to-publication. The skill-packaging pattern is the interesting bit — relevant to how Hermes’s tool surface could be versioned and shared.
❤ The Wedding Countdown
408 days
until Saturday, November 27, 2027 — the lodge. From the record: Rip City Grill still holds the food-cart crown at $14/person, weddings in writing. Vendor quotes not yet requested — plenty of runway at 408 days out.