Morning
Tech Brief

The edition where 816 billion parameters just went open-source, Intel squeezes two GPUs onto one card, and somebody rewrote Arch’s package manager in Zig.

Top signalMeta’s Muse Spark Open — 816B open weights, 42B active, 1M context. The biggest open release of the year.

On sale todayRTX Spark N1X laptops land with up to 128GB unified memory — a 70B model fits on a laptop.

Set the alarmNo Rest for the Wicked’s class-system showcase streams today at 8 AM PDT.

Beat 01

AI tools

4 stories

Meta drops Muse Spark Open: 816B of open weights for coding and agents

Announced Oct 15, weights on Hugging Face the same day: a 816B-total / 42B-active mixture-of-experts (256 experts, 8 per token), 1M-token context, text/image/video/audio in, text out. Meta’s own card has it leading the open field on tool-calling (MCP-Atlas 89.2) but trailing Kimi K3 and GLM 5.3 on long coding sessions (Terminal-Bench 2.1: 76.4 vs 85.0; DeepSWE 1.1: 60 vs 68) — a tool-calling brain that tires before the code does. It runs on AMD, Intel, and NVIDIA hardware via vLLM, SGLang, and llama.cpp, and on two M5 Ultra Mac Studios over MLX (up to 1TB unified memory). Intel even ships day-0 support on four 160GB Crescent Island D-Series cards. The license is Meta’s own Muse Model License: commercial use OK, but hosted services over $20M revenue need a separate deal. 816B is not for your 2×16GB box — but the calculator in Beat 02 shows exactly how far your hardware gets you.

Sources: Subagentic (model-card teardown), Gadgets360 (Apple partnership), benchmark comparison vs GLM 5.3 & Kimi K3.

NVIDIA’s Nemotron 3.5 trio: a 120B agent brain and two locals-sized models

Also Oct 15: Super VL, a 120B-total / 12B-active MoE with 1M context built for always-on agents (text+image+video in one checkpoint, NVIDIA claims 5× single-GPU throughput vs open peers, OpenMDW-1.1 license); Bolt 74B-A9B, a 74B-parameter agent model that fits in 24GB — that’s inside your dual-A770 budget with room for KV; and Bolt E4B, a 4GB model for voice and on-device chat. NVIDIA’s reported numbers are strong on GPQA-Diamond (89.90) and shaky on tau3-bench banking (17.78) — vendor numbers, trust-but-verify. The Bolt 74B is the one to actually try: 9B active per token, open weights, and the first serious 74B-class local candidate that doesn’t need a datacenter.

Sources: MarkTechPost, AI Parabellum.

Liquid AI’s LFM2.5-24B-A2B: 2B active parameters that want your agent workloads

Oct 13, slightly older but the week’s sharpest local-model decision: 24B total, only 2B active per token, 184 tok/s on an M5 Max, and benchmarked as the tool-calling pick against Qwen3.6-35B-A3B’s coding-depth pick. That’s exactly the talker-lane question you’re sitting on (Qwen3.6-27B Q6_K): Liquid’s thesis is you should pick your 24B-class model by workload — agents vs. code — not by leaderboard average. Worth a slot in the Glimmer-vs-Qwen evaluation queue as the efficient-MOE data point.

Source: Tech Insider (launch teardown with benchmarks vs Qwen).

Vitalik gives his AI agent a “paranoid mode”: Tor, zkAPI, local LLM

Posted Oct 16: a paranoid-mode wrapper around his agent — every conversation gets its own Tor circuit, zkAPI hides payment identity, a local LLM rewrites prompts, and the runtime is sandboxed with bwrap + torsocks forcing all network access through Tor, plus random cover traffic. He admits latency is rough and cover traffic won’t survive a coordinated model-provider/Tor-entry attack. Fun as a celebrity side-project, genuinely useful as a design pattern: the local-model-as-privacy-scrubber idea is the same trick that would make an agent you let near your tools trustworthy. Morty’s gateway approval story, but for network identity.

Source: PANews.

Beat 02

Hardware

4 stories + 2 tools

RTX Spark N1X reviews: class-leading GPU, aging CPU — on sale today

Two dies in one package (MediaTek ARM CPU + Blackwell GPU, NVLink C2C between them) with up to 128GB of shared LPDDR5X, up to 96GB allocatable to the GPU — four times any previous laptop GPU. A 70B model at 4-bit (≈41GB) now loads entirely on a laptop; the big config ships 20 Cortex cores (X925 + A725) and 6,144 shaders, with a smaller 18-core/5,120-shader variant. Reviews call the GPU exceptional and the CPU cores last-gen. For the household’s local-first ambitions this is the form factor to watch: unified memory is how you dodge the VRAM ceiling entirely. Also in the launch wave: DLSS 5 and NVIDIA’s claim of 2× the efficiency of a Core Ultra 9 285H + RTX 5070 at 60W.

Source: Tech Times (review roundup, launch specs).

Intel’s Crescent Island: two GPUs, one liquid-cooled card, first Xe3P die shots

First die shots of the Xe3P-based D-Series (“Crescent Island”) on N3P — a smaller die than Battlemage’s BMG-G31, because most of the die went to AI accelerators instead of render cores. The reference D270 carries 160GB of LPDDR5X, partner configs up to 480GB, 1.5TB/s bandwidth, and 1300 TFLOPS of FP8; Intel showed a dual-GPU OAM module with integrated liquid cooling. And the flex: day-0 support for Muse Spark Open, with four 160GB cards (640GB total) hosting the 816B model. Their “easily” is doing a lot of lifting — 640GB holds FP8 weights (816GB… wait, no: 816B params × 1 byte = 816GB > 640GB). Either they’re running FP4-class quantization or “hosts” means weights-only with zero KV room. The calculator below keeps them honest. Datacenter-only, but this is 2028’s used-market dream card.

Source: WCCFTech.

Microsoft’s Maia 300 bets on power and memory, not speed

Four compute chiplets on TSMC’s 2nm with 288GB of HBM4, and the pitch is efficiency, not speed: 40% more tokens per megawatt than leading GPU-based systems. Production deployment starts 2027 as the primary inference system for Copilot, Teams, GitHub, and Foundry — Azure rental unconfirmed. The honest read: Microsoft isn’t trying to beat Rubin on raw throughput; it’s trying to make inference a utility cost. If you’re watching cloud-model economics for the local-vs-cloud routing question, tokens-per-megawatt is the metric that moves your bill.

Source: Memeburn (via SiliconANGLE).

Bolt Graphics wants $100M at a $700M valuation — for a non-AI GPU

Reuters, Oct 15: the startup is raising roughly $100M to build “Zeus,” a GPU aimed at FP64 scientific computing, chip design, and professional rendering — deliberately not AI training. First Zeus test silicon round is done, TSMC fab planned. The contrarian bet in an AI-frenzied market: everyone else chases FLOPS for transformers, Bolt builds precision hardware for physics sims and 3D artists first, gamers later. The gaming-GPU roadmap is the honest hard part (drivers, game compat), but a precision-first chip in 2026 is a genuinely fresh silhouette.

Source: Reuters.

🧮 The 816B question — what does this week’s open-weights wave actually cost in memory?

Weight bytes ≈ params × bytes/param. The “fits” test adds 25% headroom for KV cache and runtime overhead. Try Intel’s claim: Spark Open @ FP8 on 4× D270 (640GB).

Model

Machine

Quantization

Weight footprint

—

GB (weights only)

With 25% headroom

—

GB vs machine memory

Does it fit?

—

KV + overhead included

⇅ Click a column header to sort — the open-weights agentic shootout (Meta’s own scorecard)

From Meta’s Muse Spark Open model-card comparison vs open peers. Vendor-reported: independent verification pending. Kimi K3 leads long coding sessions; Spark Open leads tool-calling.

Model	MCP-Atlas (tools)	JobBench (agent)	Terminal-Bench 2.1	DeepSWE 1.1	SWE-Atlas QnA	GPQA Diamond
Muse Spark Open	89.2	61.8	76.4	60	58	89.0
Kimi K3	82.3	54.3	85.0	68	66	93.5
GLM 5.3	84.2	61.4	83.9	67	59	91.7

Terminal-Bench 2.1: the long-coding gap, visualized

Beat 03

Software

3 stories

Shelly 3.1.7 ships RLPM — a from-scratch Zig rewrite of Arch’s libalpm

Out today, Oct 16: the Arch package manager from Seafoam Labs now carries RLPM, a native Zig reimplementation of the entire ALPM workflow — repo databases, dependency resolution, signature verification, hooks, scriptlets, payload install, the lot — dispatchable alongside the original libalpm instead of wrapping pacman. It’s opt-in (shelly config set NativePackageBackend rlpm), and the authors are admirably blunt that full compatibility isn’t finished. A package manager rewriting the library the distro’s own package manager depends on is either audacious or reckless; the Landlock/seccomp sandboxing and sealed-memfd archives suggest they know which one it could be. CachyOS made Shelly its default GUI package manager in April, so this experiment now ships to a very large audience.

Sources: Linux Compatible, Linuxiac (AUR isolation workflow, Atoll scan integration).

Faugus Launcher 2.5.0: nav sounds, lighter deps, monthly cadence

The one-person Linux launcher for Windows games (GOG, Epic, Battle.net, the random .exe in your downloads folder) keeps its monthly rhythm: 2.5.0 adds interface navigation sounds, a reworked Prevent Sleep mode, and drops a dependency — plus an option to strip covers and icons for low-bandwidth rigs. In a market with Lutris, Heroic, and Bottles, Faugus’s pitch is “simple and lightweight,” and the GTK4 rewrite in July is what put it on the map. Nothing earth-shaking; the interesting bit is the cadence itself — monthly releases from a solo maintainer is how small tools stay alive.

Source: Linux Compatible.

Release roundup: Fedora 45 slips, Rust patches fast, VS Code makes Copilot the default harness

Fedora 45 is a NO-GO — ship date slides to Oct 27. Rust 1.99.1 fixes three critical bugs in a record two-week turnaround. VS Code 1.142 makes Copilot the default agent harness (the editor wars continue to be about who owns the agent loop). Godot 4.8 Beta 1 enters feature freeze. Ubuntu 26.10 “Stonking Stingray” ships with a Rust core and a slimmer boot. Kernel 7.2.11 / 6.18.57 / 6.12.113 arrive with XFS, WiFi, and router security fixes. All from this morning’s Linux Compatible archive.

Source: Linux Compatible — October 2026 archive.

Beat 04

Games

3 stories

No Rest for the Wicked’s class-system showcase streams today, 8 AM PDT

Moon Studios’ “Wicked Inside” stream lands today at 11 PM GMT+8 / 8 AM PDT / 5 PM CEST, on YouTube and Twitch: the deep dive on the class system — 24 classes, 280+ traits, six trait slots, six factions — that arrives with the December open beta ahead of the March 2027 1.0 launch. Early Access since April 2024; the 1.0 delay was explicitly for performance and the class system. Founder’s Pack rewards still on the table if you buy before Jan 1, 2027; PlayStation pre-orders open in January. Set the alarm or catch the VOD — this is the game’s biggest systems reveal yet.

Source: Nextarca.

Beyond the Dark: Nightwatch — two-player asymmetric horror, out Oct 22 on Switch 2

Q-Games is publishing Atlantis Studio’s co-op horror in Asia with an Oct 22 Switch 2 launch: one player explores the haunted house in first person, the other plays Overseer on the cameras, guiding their partner out before the Gloam gets them. Couch co-op, GameShare (works across Switch 2 and original Switch), local network, and online. Troy Baker voices it. The asymmetric-information design — each player sees only part of the picture — is the co-op subgenre’s strongest trick; keep this one in mind for a couch night.

Source: Business Intelligence (Media OutReach).

Thursday’s shipping batch: DayZ Badlands, Frosthaven 1.0, CK3’s Khans

WorthPlaying’s daily roundup: DayZ’s Badlands expansion is out now, Frosthaven hits 1.0 on PC, Crusader Kings III gets the Khans of the Steppe expansion + Coronations event pack, Gravity Circuit 2 has a playable demo, Quarantine Zone’s Miami Zero update is free, and Warframe opens Epoch Operation: Taubound. A healthy Thursday for the backlog — the DayZ and Frosthaven launches are the ones with legs.

Source: WorthPlaying.

Beat 05

GitHub repos

4 repos

⭐ Click a column header to sort — this week’s picks for agent harnesses & agent skills

Drawn from this week’s trending digests and star-velocity lists. Stars are snapshots from the digests, not live counts.

Repository	Lang	Stars	Why it’s here
CopilotKit/openbot	TS	≈2k in 4 days	Open-source AI coworkers where each agent gets its own computer — real browser, own files, granted tools only — with a gateway that decides every action before it happens and records it after. AG-UI based, MIT, Docker Compose. The supervised-computer model is the architecture your phone-console breakfix work has been circling: read the gateway design.
code-yeongyu/lazycodex	TS	≈3.7k	“LazyVim for Codex”: an agent harness for complex codebases with hierarchical project memory, decision-complete planning, durable execution, and verified completion for open-ended tasks — plus selectable sub-agent roles. The evidence-over-vibes completion loop is the piece every home harness (yours included) should steal.
movomoto/better-harness	TS	≈2.4k	Runs through your coding agent and turns project and session evidence into prioritized improvements with verifiable next steps — and keeps missing evidence explicit instead of papering over it. MIT. The harness-improvement loop as a repo: delegate the loop, improve the loop.
jrenaldi79/harness-engineering	n/a	fresh	Harness engineering = context engineering applied to coding agents: rule files that load only when relevant, git hooks and linters doing the enforcing that AGENTS.md can’t (Karpathy-endorsed philosophy), planning before building, docs kept in sync. A field guide with CLAUDE.md templates — bootstrap in one command.

❤ The Wedding Countdown

407 days

until Saturday, November 27, 2027 — the lodge.
From the record: the whole weekend is booked — Fri Nov 26 through Sun Nov 28, 2027, under $10k for everything. Vendor quotes not yet requested; Rip City Grill still holds the food-cart crown at $14/person.