DeepSeek went open. Intel bet on memory.

NVIDIA · Released Oct 15, 2026 your queue

NVIDIA shipped Nemotron 3.5 Bolt: 74B of MoE stuffed into 24GB at 2-bit

This is the one you said you'd download yourself, Matt. 74B total, 9B active per token, quantization-aware-trained at 2-bit so the whole thing packs into 24GB — and it runs at launch through llama.cpp, vLLM, Ollama, Unsloth, LM Studio, and "directly in Hermes and OpenClaw." NVIDIA's own claim: faster than popular dense models, smarter than most MoE models at a similar memory footprint. Weights, data, and recipes open like other Nemotron releases. The sibling to note: a 4GB Bolt E4B for game NPCs lands Oct 23. Your homework writes itself — this is the natural challenger in the Qwen spec-pair's neighborhood, though the 24GB footprint means it needs your whole 32GB pool, no parking it on one card.

DeepSeek · Open-sourced Oct 20, 2026 day's #1 repo

DeepSeek open-sourced DeepInfer — the actual production engine behind its cheap API

Not weights. The plumbing. FlashMLA, DeepGEMM, DeepEP kernels, MoE expert parallelism with load balancing, DSpark confidence-scheduled speculative decoding, batch-invariant deterministic inference — Apache 2.0, ~9.4k stars on day one. But be honest about the catch: it serves one model (V4.1-Flash) on two chips (NVIDIA SM100, Huawei Ascend 950), and the load-balancer machinery that makes the economics work stayed home. Nothing here runs on your A770s. What does transfer: DSpark-style speculative decoding and FlashMLA attention kernels are the exact ideas the llama.cpp folks reverse-engineer next. Worth a star, not a weekend.

Google · Open-sourced Oct 20, 2026

Google open-sourced Harness²: a better agent harness, weights left untouched

Frozen weights, warmer scores — your exact doctrine, Matt. Google's play: improve the harness around the model instead of retraining the model, and Gemini 3.5 Flash gained 8.4 and 7.5 points on two task benchmarks just from that. This is the kind of free lunch the hermes-desktop workstream should be stealing ideas from — tool surface and orchestration are where the agentic gains live once the brain is good enough. Skim the harness diff before the next "let's swap the model" impulse.

Moonshot AI · Released Oct 14, 2026

Kimi Code Desktop 1.1.0 lands on Linux — with Remote Control to steer the agent from your phone

The phone-steers-local-agent dream, productized. AppImage/deb/RPM builds, a file-tree panel, and Remote Control: watch task progress, approve things, and send follow-ups from a phone, tablet, or another machine via link/QR code while files, shell, and tools keep running on the desktop. That's the interaction shape the hermuse phone console is chasing — difference being Kimi did the pairing link in one release. Study the UX, steal shamelessly, don't get sad about it.

Industry trend · October 2026

The decision-model wave: Cloudflare Clef (27B/9B) and Amazon's free local Strands Decider 2B

Typed answers, probabilities, no prose — the kev-decision playbook, now an industry pattern. Cloudflare's Clef models are post-trained from Qwen bases, read text/JSON/images/video, and cost $0.24/$0.09 per million input tokens with 65k context. Amazon's Strands Decider is 2B and free to run locally. The point of the pattern: put a cheap check right before the action instead of paying for a frontier round-trip. That's literally what :8080 is for. Meanwhile in megamodel land: Meta's Muse Spark Open (816B/42B active, 1M context, limited-commercial license) and Reflection's Beam (501B MoE, Apache 2.0, coding/agentic, NVFP4 squashes 1TB → 280GB) are datacenter toys. Admire from afar.

NVIDIA · Shipping since Oct 16, 2026

RTX Spark laptops are actually shipping — unified memory is no longer a keynote slide

128GB of one shared pool is what your 32GB split-pool setup dreams about at night. Eight OEMs, starting at $2,599.99 for Microsoft's entry Surface Laptop Ultra, up to ~$6,999.99 for ASUS's fully loaded ProArt P16 (128GB/2TB). NVIDIA claims 120B-param models with 1M context on-device — unverified, treat accordingly. Early reviews report real tradeoffs: AI throughput ≈ midrange discrete laptop GPU, the Arm Grace CPU trails M5 on single-core, emulated x86 takes a hit, and battery costs you 20–25% AI performance. Compact desktops land in November. The honest read: Spark is the unified-memory endgame arriving pre-built, but $2,599 buys a lot of patience for the DIY route — your pooled VRAM isn't obsolete, it's just unfashionable.

Intel · OCP Global Summit, Oct 12–13, 2026

Intel's "Crescent Island" is real: GPU D-Series D270, 480GB LPDDR5X, 1.3 PFLOPS FP8

Intel is betting the whole company on your thesis: inference cares about memory capacity, not training flops. Xe3P architecture, 1.5 TB/s bandwidth, 350W air-cooled PCIe card — one card handles 300B-param models, eight handle a trillion+. Partner cards go to 480GB of LPDDR5X (Intel's own card: 160GB). The engineering trade is the story: LPDDR5X over HBM means less peak bandwidth but lower power, cost, and supply exposure — and it drops into existing data centers without special racks. Sampling to customers now, no launch date, no price. The dragon your hardware.py rewrite is chasing has an official name now.

NVIDIA · Launching Oct 23, 2026

64GB DGX Spark ships Friday — the cheaper on-ramp to unified memory

NVIDIA noticed people choking on the 128GB price tag. The 64GB tier supports up to 100B-param models under optimized quantization (per NVIDIA), and multi-unit desktop clustering is on the menu for scaling unified memory. Ships through Acer, ASUS, Dell, Gigabyte, HP, MSI. Still enterprise-shaped, still not your box — but it's the clearest signal yet that unified-memory local inference is a real product category now, not a demo.

Interactive · memory pools, unified and otherwise

Unified memory pool comparison — where your rig sits in this week's landscape

The teal bar is not small because it's wrong — it's small because the others are absurd. Your 32GB pool is the scrappy underdog that actually runs on a 9800X3D box without a data center's blessing. Vendor max-param claims are self-reported; treat all four bars as marketing with error bars.

Interactive · VRAM math, your 92% rule edition

Will it fit your rig? — the model-fit calculator

Sanity-check Bolt before you download it, Matt: slide to 74 params × 2-bit and see what 24GB of weights plus KV growth does to a 32GB pool.

Omarchy ecosystem · Released this week your #1 item

Strata 1.0 is out — the new Linux file manager built for Omarchy

Hermes's priority #1 just made the Linuxiac weekly wrap. Miller columns, previews, the whole package. The only thing this release needs is the thing that was always going to be the decider: a screenshot of Miller columns + previews, run on your actual box, proving the install. Rick has his eyes on this one — the public release existing is nice; the exercised proof is what closes it.

Arch packaging · v3.1.7

Shelly 3.1.7 ships a new Zig-based engine — the RLPM you shelved, arriving upstream

Fate has a sense of humor: you parked RLPM, and now the ecosystem is shipping the Zig backend itself. Watch whether Shelly's engine lands the parallelism and speed story cleanly — if it does, that's the engine you would have written, and your watch-list item updates from "build" to "adopt or audit." Either way, no building before the shootout queue is clear.

Distro & desktop · Week 42

Omarchy 4.0.5 (kernel 7.2.8); KDE Plasma 6.8 drops X11 after 30 years; GNOME 51.1; Firefox 158 gets JPEG XL

Pour one out for X11 — thirty years of service, retired by the KDE team. Your Wayland-native stack shrugs at this, which is the point: you made the right migration call years early. Omarchy 4.0.5 brings kernel 7.2.8 with better hardware support (a quick pacman -Syu story, not a project). Firefox 158 finally shipping JPEG XL is the quiet win — faster-loading images everywhere with no plugin tax. Honorable mentions from the same wrap: SparkyOS 2026.10 (portable Arch on a USB stick, persistent — the whole OS as a thumb drive), Libreboot 26.10 with full UEFI support and a brand-new anti-AI clause in its AGENTS.md (the irony of an AI policy written for the robots is not lost), and BeamFS — an experimental self-healing filesystem — proposed for kernel inclusion.

BSD · Released Oct 20, 2026

OpenBSD 8.0: the first BSD release that speaks your GPUs' language

inteldrm now supports Intel Arc A-Series (DG2/Alchemist) — your cards, on OpenBSD. The drm stack is rebased on Linux 6.18, and amd64's kernel virtual address space jumped from 4G to 512G — which now allows mapping more than 4GB of VRAM. That's a hardware.py-adjacent nugget hiding in a BSD release note. You're not switching to OpenBSD, obviously — but if the BSD people are mapping 4GB+ of Arc VRAM before your kernel config got there, it's a good mirror to hold up.

ID@Xbox × IGN Showcase · Oct 20, 2026

Indie avalanche: Mouthwashing finally escapes the PC, Shrek 2 gets remastered, and 2027 fills up

Mouthwashing hits consoles Oct 29 — the horror backlog grows, and it was always the most disturbing thing on your "maybe later" list. Two years after its PC debut, Wrong Organ's space-freighter nightmare arrives on Xbox. Around it: Shrek 2 Remastered (Nov 10, 2004 film tie-in resurrected), The Wolf Among Us Remastered (Oct 29), Scramble Knights Online (Oct 26), and A Game About Digging a Hole Together (Dec 4). Critical Reflex landed three ports in one stretch. The indie slate is healthy; your playtime budget is not.

Xbox Game Pass · Wave 2 announced

Game Pass Wave 2: FFVII Remake Intergrade lands, plus two day-one releases

Final Fantasy VII Remake Intergrade hitting Game Pass (Oct 29, all tiers) is the steal of the month. Day-one additions: Scramble Knights Online (Oct 26, also above) and Terrible Lizards (Oct 29). Sandustry drops today (Oct 21) in Game Preview on PC. Pit of Goblin rounds it out Nov 3. If your A770 has been idle for actual games — and the gaming GPU rule says slot 1 is for gaming — Intergrade is the excuse to warm it up.

PlayStation Plus · October catalog

PS Plus October: Mafia: The Old Country, Horizon Forbidden West returns, Mycopunk day-one

Mycopunk as a day-one catalog addition is the swing; Dino Crisis 2 on Premium is the nostalgia bomb. Mafia: The Old Country and a returning Horizon Forbidden West anchor the catalog (Oct 20). Monthly games side: F1 25, Hunt: Showdown 1896, EDF: World Brothers 2 (claim window Oct 6 – Nov 2). Not your platform of record, but Hunt's extraction-loop is the kind of thing that survives every hardware generation — worth knowing it went free.

Release calendar · This week

Oct 29 is a traffic jam: Phantom Blade Zero, Titanium Court's console debut, ANOMALITH — and CoD: Modern Warfare 4 Friday

Call of Duty: Modern Warfare 4 drops Oct 23 — the annual GPU-stress test your slot-1 card exists for. Then Oct 29 packs Phantom Blade Zero (the stylish action one to watch), Titanium Court escaping Steam for Switch/Switch 2/PS5/mobile (IGF Grand Prize winner, match-three-meets-RPG — genuinely one of the year's best), and ANOMALITH, a RE-inspired survival-horror dropping two days before Halloween. October's fourth week is doing a lot of work.

Velocity measured 2026-10-21 via ghtrends.dev official star-history data

Already on your stack: mattpocock/skills (#2 weekly Oct 7–14 — now installed) and cathrynlavery/diagram-design (#10 — trial on the local pair). You adopted the winners before the crowd caught up.