The Morning Brief No. 55

MiniMax releases open weights for 428B-parameter M3.1 Flash

What: Weights out Oct 23 for a 428B-total / 23B-active mixture-of-experts model, 1M-token context, text + image + video input, built for coding, agents and computer-use — with SGLang/vLLM deployment guidance.

Take: Read the small number, not the big one. 23B active is the part that actually runs per token — but the full 428B still has to live somewhere, so at Q4 that's ~214 GB of weights (see the table below). This is not a dual-A770 living-room download; it's a datacentre/API model you can now inspect, distil from, and run on rented iron. The honest home-lab win is the architecture signal, not a new daily driver.

TechNode, Oct 23 · MiniMax M3.1 Flash

Kimi K3.1 Pro: 3 points from the frontier, weights promised by Nov 4

What: Moonshot's K3.1 Pro launched Oct 23 (Kimi, Kimi Work, Kimi Code, API at $2 / $10 per million input/output tokens). Same 2.8T-param / 104B-active base as K3, with much larger post-training compute. Artificial Analysis scores it 55.1 — 4th overall, behind Claude Opus 5.5 (57.6), Sonnet 5.5 (56.0) and Muse Wonder Full (55.8) — and 67% on Terminal-Bench 4.0 vs 13% for K3.

Take: The number that matters for Matt's stack: Terminal-Bench 67% is a genuine agentic-coding leap, not a vibes bump. The catch is in the headline — “open-weight” is still a promise, not a download, and Moonshot deliberately trained some cyber capability back out before release. Calendar Nov 4; judge it when the weights and licence actually land.

Artificial Analysis Intelligence Index, as reported Oct 24. Bars scaled to the leader.

CellCog / Artificial Analysis, Oct 23–24 · also WebProNews on the cyber-data removal

A distribution week: Agent Skills go to the Linux Foundation, OpenAI & Claude land on Google Cloud

What: Oct 22: Anthropic donated the Agent Skills open standard to the Linux Foundation-directed Agentic AI Foundation (70+ platforms, ~1M skills, ~280M installs per Anthropic). Oct 21–22: OpenAI models enter Google Cloud's Model Garden and Claude on GCP goes GA. OpenAI's Decisions API also went GA — GPT-6 Luna for bounded probability/choice/score answers, ~$0.10/M input tokens, no output-token charge.

Take: Boring headline, important plumbing. Skills just went from “Anthropic's format” to foundation-governed infrastructure — which is exactly the direction Matt's Hermes skill stack wants. And a decisions API priced like that is a standing invitation to stop burning a frontier model on yes/no questions. (Yes, that's the d1/Kev lane, just priced by someone else.)

DEV roundup of the Oct 19–25 week, published Oct 24

TanStack AI 1.0 ships: stable agent APIs, 23 provider adapters, MCP support

What: The toolkit that started as a single chat() function is now a stable 1.0: chat, tools, media, MCP, persistence, sandboxes and agents, with 23 provider adapters and 8 UI frameworks, built by 140+ contributors.

Take: “No lock-in, ever” as a formal commitment is the whole pitch. If you're wiring agent features into a web app — Hermuse-adjacent thinking — this is the boring, portable runtime you actually want, not another framework that marries you to one provider's SDK.

GenZ NewZ, Oct 23

Nvidia's Vera CPU wins Modal — “2× agent throughput” under load

What: Modal, the AI cloud startup that previously ran only Intel/AMD, will start hosting agents on Nvidia's Arm-based Vera CPU. Modal's early tests: up to 2× task throughput vs Intel Granite Rapids, and (per WccfTech's read of the same tests) over 6× vs AMD EPYC at 150 concurrent containers — credited to high memory bandwidth holding up under contention. Modal raised $750M at a $15.75B valuation; Nvidia targets $20B+ in Vera sales this fiscal year.

Take: The 2× is Modal's own benchmark against ~2-year-old rival chips, not an independent audit — and Modal is keeping Intel/AMD in a mixed fleet, not divorcing them. Still, the shape of the claim is the interesting part: Vera wins when the box is crowded, not when it's idle. That's an agent-sandbox story, not a gaming story, and it maps uncomfortably well onto how Hermes-style workloads actually behave.

Reuters, Oct 19 · WccfTech / Fudzilla follow-ups

Steam survey, September: AMD at 48.4% of CPUs, Nvidia at 76% of GPUs

What: Intel's CPU share fell ~1.9 points in a month to 51.6% vs AMD's 48.4% — the crossover point is in sight. Nvidia GPUs climbed 3.1 points to 76%; the fastest-growing card is the RTX 5070 (6.2% of surveyed systems). Octa-core CPUs hit ~30%.

Take: Matt's 9800X3D is, once again, on the right side of history. The survey is a lagging indicator of what gamers actually bought, and right now they're buying X3D cores and mid-range Nvidia cards — presumably while muttering about AI pricing, like everyone else.

Digital Foundry / Steam Hardware Survey, Oct 2 (September data)

OBS Studio 33.0: CEF jumps 127 → 150, plugin loading overhauled

What: Released Oct 23. Chromium Embedded Framework leaps from July 2024 to September 2026 tech; first-party plugins move to a new core folder and third-party plugins get a revised install location. Legacy plugin locations keep working — until OBS 34.0, when they stop. Also: better source snapping, NVENC on Arm64, and Linux fixes for USB-mic disconnect crashes and duplicated PipeWire screencasts.

Take: The CEF jump is the real story — browser sources stop living two years in the past. The plugin move is a polite countdown: if you run third-party OBS plugins, migrate them now, not the week 34.0 ships and your scene breaks live.

Linuxiac / Phoronix, Oct 23 · official release notes

Bazzite 44.20261023: weekly build lands, two new handhelds enter testing

What: The Fedora-Atomic gaming distro's Oct 23 stable build brings Gamescope 3.16.31, InputPlumber 0.83.0 and a fresh UMU Launcher — and opens public testing for the ASUS ROG Xbox Ally X20 and Acer Predator Atlas 8.

Take: Bazzite's “never missed a weekly build in months” cadence is the quiet flex here. Handheld testing pools are where SteamOS-likes actually get good — if either of those handhelds is on your radar, the testing builds are where the bugs get found, by volunteers, on purpose.

LinuxCompatible, Oct 23

Self-Host Weekly: PHP2SHELL leaked early — patch PHP apps ASAP

What: Two critical PHP vulnerabilities (PHP2SHELL) leaked ahead of their planned patches, PHP rushed the fixes out, and the weekly's advice is blunt: update PHP-based apps (and Unraid) immediately. Same issue: OpenBSD 8.0 shipped (new GPU support, ARM64 hibernation), Linus Torvalds made his first known AI-assisted kernel contribution, and Plex reportedly laid off its player dev team.

Take: If anything on the home lab runs PHP — and something always does — this is the weekend's five-minute job. The Torvalds footnote is the fun one: the kernel's maintainer using AI assistance is no longer a hypothetical debate, it's a commit.

Self-Host Weekly, Oct 23

Also out Oct 23: OpenSSL 4.1, LibreSSL 4.4, Calibre 9.17, Shelly 3.1.9

What: OpenSSL 4.1 (more AVX-512 optimisations), LibreSSL 4.4 (post-quantum TLS improvements), Calibre 9.17 (JPEG XL support, faster MOBI/AZW3 conversion), and Shelly 3.1.9 — the Arch package manager Matt watches — with file-migration support and AUR upgrade fixes. Plus oneDNN 3.14 gains ACE support and early AMD Zen optimisations.

Take: The Arch-adjacent one is Shelly, obviously — file migration plus AUR fixes is exactly the kind of unglamorous reliability work that keeps a rolling system boring. oneDNN's Zen work is the sleeper for local inference on AMD silicon.

Linuxiac / Phoronix, Oct 23

Call of Duty: Modern Warfare 4 launched Oct 23 — DMZ is back, $69.99, not on Game Pass at launch

What: Campaign, multiplayer and the returning DMZ extraction mode, on Xbox Series X|S, PC, PS5 and Switch 2. Campaign early access ran from Oct 16 for preorders; the full game is not included with Game Pass at launch.

Take: DMZ returning is the actual news — extraction modes have quietly become CoD's most interesting corner, and this one launches straight into a crowded shooter season. Remember the house rule before launch night: games run on the slot-1 A770… well, in spirit. (This one's a PC/console title, not a lab benchmark — the GPU rule still applies to whatever you actually play it on.)

G2A / Tech-Gaming release listings, Oct 23

Game Pass Wave 2: Final Fantasy VII Remake Intergrade headlines Oct 29

What: Wave 2 (Oct 21 – Nov 3): Sandustry (Oct 21), Dodo Duckie (Oct 22), The Séance of Blake Manor (Oct 27), then Oct 29 brings FFVII Remake Intergrade, Plants vs. Zombies: Replanted and Terrible Lizards, with Pit of Goblin closing Nov 3.

Take: Oct 29 is the date to circle — FFVII Remake on Game Pass is the kind of “finally, no excuse” drop that eats a weekend whole. Séance of Blake Manor on the 27th is the spooky-season chaser, timed suspiciously well for the week before Halloween.

Tech Insider / Xbox Wire reporting, Oct 23

This week's shelf: FINAL FANTASY RESONANCE, PEAK on console, and GTA VI in 26 days

What: Oct 22–28 also brings FINAL FANTASY RESONANCE (HD-2D, $49.99), PEAK to console ($19.99), Warrior Cats: Clans of the Forest, SULFUR and Cotton Reboot!. And the calendar beyond: Grand Theft Auto VI is dated Nov 19 — 26 days from today.

Take: RESONANCE is the critic's pick of the pile; PEAK on console is the couch-co-op pick. But let's be honest — everything in this beat is now a countdown to Nov 19, and every other release is just something to play while waiting.

Tech-Gaming new releases, Oct 23