The Morning Brief — Edition No. 50

🧠AI tools

open weights season — the heaviest drop week in a while

NVIDIA Nemotron 3.5: three open models, one aimed straight at your box

The one you've been waiting on since the manual-download plan: a 120B multimodal Super VL (12B active/token) for always-on agents, plus two Bolt models for local deployment — including the 74B-A9B native-2-bit GGUF you asked for the link to. NVIDIA claims 5× throughput of comparable open models on agent work (their numbers, treat as marketing). OpenMDW-1.1 license — not OSI-open — and it needs recent llama.cpp with nemotron_h_moe support. The third file for spec-decoding still lives in the separate LiLiCorr drafter repo.

GenZ NewZ — Oct 15 (models announced Oct 15)

Reflection AI ships Beam weights under Apache 2.0

501B total, 23B active sparse MoE, genuinely Apache 2.0, day-zero vLLM support. This is the anti-Meta play: no attribution clause, no revenue-triggered relicensing — the exact shape your genuinely-open-licensing filter has been waiting for. Caveat: the release page ships no benchmark scores (their efficiency claims are their own), so Beam stays a candidate until someone independent runs it. At 23B active it's a monster for the box though — Q4 alone is ~300GB, so this one's a file-server dream, not a 32GB-box reality.

OpenDataScience weekly roundup — Oct 18 (weights released Oct 16)

Meta releases Muse Spark Open: 816B params, custom license

The biggest drop of the week and the one with the most strings: 816B total / 42B active, 1M-token context, multimodal — behind a "Muse Model License" that demands "Built with Muse Spark Open" attribution and a side deal with Meta if your API revenue clears $20M. Meta even tuned it with Apple so it runs on clustered M5 Ultra Mac Studios (two 512GB boxes for the 4-bit version). Open-ish, agent-weak (trails Kimi K3 on OSWorld Verified 84.8 vs 80.4), and the license keeps it off any box you care about. Fun trivia: I'm literally running on Muse Spark right now — the closed one, not this.

OpenDataScience weekly roundup — Oct 18 (released Oct 15)

DeepSeek V4.1 Flash + Search/Fetch API: 77% on DeepSWE at $0.017/M tokens

Quietly the best value play of the week: 77.0% on the DeepSWE benchmark at an effective $0.017 per million tokens, plus a new Search and Fetch API for live context. That's cloud-chaff pricing for a model that can write real code — a natural fallback for the desktop app's cloud lane, or the thing Hermes uses when the local pair is busy. Also in the same 60 seconds of news: Anthropic floating federal AI legislation to Congress, and Anthropic formalizing an independent eval deal with METR.

BushAIcave daily digest — Oct 18

This week's weight drops — VRAM footprint at Q4_K_M (4.85 bits/param)

Muse Spark Open 816B

≈494 GB

Reflection Beam 501B

≈303 GB

Nemotron 3.5 Super VL 120B

≈73 GB

Nemotron 3.5 Bolt 74B

≈44.9 GB

Bolt 74B at native 2-bit

≈23.7 GB ✓

Liquid LFM-2.5-24B-A2B

≈14.5 GB

Your Kev-4B (scale ref)

≈2.4 GB

Weight-only estimate: params × 4.85 ÷ 8. Bolt's native 2-bit (2.56 bpp) is the only headline model that physically fits the dual-A770 pool at Q4 — everything else above it needs a server or a loan.

Does it fit your box?

Params × bits-per-param ÷ 8, plus KV/context and runtime overhead. Your standing rule: base + max KV growth must clear 29.4GB (92% of the 32GB pool) with margin.

Parameters (B)74 Quantization KV/context + overhead15%

0 GB

Bar scaled to 32GB pooled dual-A770 · red line = your 29.4GB (92%) standing rule · a single A770 = 16GB

💽Hardware

the "agent PC" category is now a real product line

NVIDIA RTX Spark laptops ship: Blackwell + Grace, 128GB unified, from $2,599

ASUS, Dell, HP, Lenovo, Microsoft and MSI are rolling out RTX Spark Windows machines today: a Blackwell RTX GPU plus a Grace CPU co-designed with MediaTek, up to 128GB of unified memory and 1 petaflop of FP4. NVIDIA is explicitly pitching them as local agent PCs — and they ship Nemotron 3.5 Bolt 74B tuned for them. The sleeper feature: NV PAIR, which lets the Spark box serve inference to other PCs on the home network. Translation: this could be your always-on inference node while the PC stays free for the fun stuff. $2,599 starting is real money, but the category is now undeniable.

IT-Online — Oct 19

Intel's GPU D-series: 480GB of LPDDR5X on one card (the Crescent Island reveal)

Intel's first D-series accelerator (Xe3P, ex-Crescent Island): up to 480GB of LPDDR5X, 1.5TB/s bandwidth, 1.3 petaflops FP8, 350W air-cooled PCIe — no HBM, no ray tracing, no gaming pretense. Intel's thesis: inference is a capacity game, and laptop RAM is the cheat code. One card holds ~300B params; eight cards pass a trillion. As an Arc owner you can appreciate the poetry — Team Blue's GPU division went full data center while you're still coaxing SYCL out of the A770s. Watch it for where the industry's memory economics are going, not as a shopping list.

AbsoluteGeeks — Oct 19 (shown at OCP Summit Oct 12)

Mechrevo iMini P700 gets AMD: Zen 4 8C/16T, up to 128GB RAM in a sub-liter box

Today's mini-PC drop: 8-core/16-thread Zen 4 with Radeon 780M, dual SODIMM slots taking up to 128GB, dual 2.5G Ethernet. No discrete GPU, so it's a CPU-inference / homelab node shape — 128GB of RAM in something the size of a sandwich is a lot of llama.cpp headroom for cheap. No pricing yet. Given your N95 mini-PC lineage, this is the "what would the next ghost box look like" spec sheet.

GizNewsDaily (via ITHome) — Oct 19

🐧Software

kernel day, plus the most unhinged package-manager move of the year

Linux 7.3 released today: FailFS, scheduler changes, Zen 6 groundwork

Fresh off Torvalds's Sunday release ("this week I somehow kept my wits together long enough to keep track of the day of the week"): new FailFS, scheduler changes, better memory handling, FUSE zero-copy I/O via io_uring, AMD Zen 6 prep, graphics-memory improvements for low-VRAM AMD GPUs, and more Rust abstractions. Arch will pick it up any day — and the 7.4 merge window opens today, with 7.4 likely becoming this year's LTS. Note for your stack: your omarchy kernel is its own package, so omarchy-update won't pull 7.3 until the linux-omarchy build lands — don't expect Binder to appear in it either.

Linuxiac — Oct 19

Omarchy 4.0.5: kernel 7.2.8, better hardware support

Your distro shipped a point release with kernel 7.2.8 and hardware-support fixes. Small but it's yours — run the update when convenient. Same week in Omarchy-land: Strata 1.0, a new Rust/GTK4 file manager built specifically for Omarchy — Miller columns, Vim-style navigation, bubblewrap-sandboxed previews, SMB/SFTP via GVfs. A file manager that speaks fluent Vim is very on-brand for your setup; worth a look as a Thunar/Dolphin alternative.

Linuxiac weekly wrap-up — Oct 18

Shelly 3.1.7 ships RLPM: a from-scratch Zig rewrite of Arch's libalpm

The Arch GUI package manager just grew a second brain: RLPM, a native Zig reimplementation of the library pacman itself depends on — repo reads, dep resolution, archive verification, hooks, scriptlets, the full ALPM workflow, hand-rolled. The authors are admirably honest that it's not finished. Rewriting the package core of your own distro's ecosystem in Zig is either madness or the most Arch thing ever committed to git — possibly both. Seafoam Labs's release cadence (nearly every few days since August) suggests you should let this bake before trusting your system to the parallel backend.

LinuxCompatible — Oct 16

KDE Plasma 6.8 drops X11 sessions — after 30 years of KDE

Three decades in, KDE is finally deleting X11 sessions. The long goodbye to X is nearly complete on the desktop side; if you're still running anything X-dependent on the box, the clock is audible now. Your Hyprland/Omarchy life is Wayland-native already, so this is a spectator sport — but it's a genuine end-of-era beat. Also this week: GNOME 51.1, systemd v262.1, OBS Studio 33.0.0-RC2, and SteamOS 3.9.3 beta fixes.

Linuxiac weekly wrap-up — Oct 18

🎮Games

preservation wins and final patches

Oblivion gets a surprise PC update nearly 20 years on

GOG rolled out a refreshed 2006 build of The Elder Scrolls IV: Oblivion — stability patches for modern multi-core CPUs, controller support, and improved rendering baked straight into the installer, no community patch stack required. The preservation angle is the real story: the classics people actually bought are rotting on modern hardware, and storefronts are finally doing the maintenance themselves. Analysts expect more mid-2000s titles to get the treatment. Your backlog called; it says you never finished the Dark Brotherhood questline.

ADN News — Oct 18

Beast of Reincarnation: today's patch is the last major one

Game Freak says the October 19 patch ends major post-launch support for Beast of Reincarnation — nine patches in two and a half months, covering camera distances, DLSS/FSR, FOV slider, cutscene skipping, difficulty settings, and the October 9 update that let players skip three mid-game regions. No DLC or sequel announced; the studio's eyes are on Pokémon Winds and Waves (Switch 2, 2027). A remarkably honest post-launch arc: they fixed the game, then said so and walked away.

GamingProMax — Oct 18

Release roundup: Enshrouded 1.0, Frosthaven 1.0, Castlevania: Belmont's Curse, more

A stacked October 15 release slate: Enshrouded and Frosthaven both hit 1.0 on PC, Castlevania: Belmont's Curse and Valor Mortis are out, plus Tales of Eternia Remastered and Harvest Moon: Echoes of Teradea. Also announced: Drova II: Torn Lands, a sequel to the grimdark open-world action RPG. If the wedding fund ever allows a game splurge, Enshrouded 1.0 after its long early access is the one with the proven track record.

RPGamer news — Oct 15

Monster Hunter Wilds Switch 2 Prologue demo drops next week

Capcom's bringing Monster Hunter Wilds to Switch 2 with a Prologue demo next week, digital pre-orders opening alongside. The handheld-PC-to-Switch-2 porting pipeline is getting interesting — and given the Zelda recomp on your horizon, Nintendo's hardware remains the gift that keeps on giving to the tinkerer's backlog. Worth a download if only to see what Capcom squeezed out of the hardware.

RPGamer news — Oct 18

⭐GitHub repos

the internet is starring agent infrastructure at warp speed

Repo	Stars	▲ / day	License	What it does
morluto/rea	331k	+15.4k	MIT	Agent-driven reverse engineering: Ghidra/Hopper/IDA, APKs, firmware, Electron apps
Graphify-Labs/graphify	129k	+626	Apache 2.0	Codebase → queryable knowledge graph, tree-sitter local, no vector store
stablyai/orca	97.6k	—	MIT	Fleet orchestrator for coding agents: parallel worktrees, mobile companion (AUR: stably-orca-bin)
google-research/harness2	9.7k	+6.1k	Apache 2.0	Recursive agent harnessing: agents rewrite their own prompts/tools/memory
feder-cr/git-ai	2.4k	+792	MIT	Open implementation of Anthropic's Scion — git redesigned for agent fleets

morluto/rea — the #1 trending repo today, and it's your REA pursuit in a box

331k stars and +15.4k in a day. REA points your coding agent at any app — native binaries through Hopper/Ghidra/IDA, Electron apps, .NET, websites — and has it explain how a feature works with evidence, locally. This is the bounty-hunter stack you had Hermes installing, packaged as an MCP skill. The APK/Firmware guides line up with your MobSF/Exodus workstream. If anything on this list deserves a Monday-night install, it's this one.

GitHub — verified today · MIT

google-research/harness2 — agents that rewrite their own harness

Google Research's recursive harnessing framework: propose harness edits, probe them on the same task, compose the winners — weights untouched. Up to +11.5 points on WorkBuddy-Bench. The interesting bit for you: the eight editable components read like a design doc for what your Hermes desktop harness already does by hand (system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, plugins). Apache 2.0, OpenCode and Codex executors. Steal the ideas, not the code.

GitHub — verified today · Apache 2.0

Graphify-Labs/graphify — query your codebase instead of grepping it

Turns code, docs, SQL schemas, configs and PDFs into a queryable knowledge graph — deterministic tree-sitter parsing, zero LLM credits for the code pass, every edge tagged EXTRACTED or INFERRED. 0.497 recall@10 on LOCOMO. For a home-lab guy with agents crawling repos: query instead of read is the token-budget move your Hermes trajectories would love. 128k stars and a /graphify skill for every major coding assistant.

GitHub — verified today · Apache 2.0

feder-cr/git-ai — Scion, but open (created 4 days ago)

Anthropic built Scion, a server-centric VCS for hundreds of concurrent coding agents, and kept it internal. Someone just built the open version: server-side history, compare-and-swap landing instead of merge queues, changeset log, per-agent tokens. Their 50-agent benchmark: 1,862–2,544 lands/min vs 240–272 for classic git. Created October 15 and already +792/day. If you ever run agent fleets that fight over one branch, this is the plumbing they were missing.

GitHub — verified today · MIT

The Morning Brief — Edition No. 50

🧠AI tools

This week's weight drops — VRAM footprint at Q4_K_M (4.85 bits/param)

Does it fit your box?

💽Hardware

🐧Software

🎮Games

⭐GitHub repos

The Wedding Countdown