Morning Tech Brief — October 13, 2026

BEAT 01

AI Tools

4 stories

Meta launches Muse Wonder & Wonder Full — and goes closed-weight

Meta announced its new frontier pair on Oct 12: Muse Wonder (Model API preview at $4/$20 per million input/output tokens) and the larger Muse Wonder Full, exclusive to Muse Code subscribers ($50 or $200/month tiers). The real story is the business model: both are closed models served via API — a departure from the open-weight Llama lineage — with Meta's AI chief Alexandr Wang pricing them "meaningfully below" OpenAI and Anthropic flagships. Internal codename was reportedly "Watermelon." Caveat: pricing/benchmark figures come from Meta-briefed reporters; no independent verification yet. Source: TrendInTech.

The headline isn't the model, it's the pivot: the house that open-sourced Llama is now selling closed weights by subscription. Undercutting the flagships on price is easy to claim pre-verification — watch the independent benches. Also, your assistant now shares a first name with Meta's frontier model, which is going to make channel mentions confusing.

llama.cpp grows a native System One decision-model endpoint

llama.cpp now supports local decision-model inference through a System One endpoint that scores a fixed set of choices in a single forward pass — working on small CPU-friendly models and accepting some image inputs. Dated ~Oct 10. This is directly on-target for a Kev-4B-style local decision stack: single-pass choice scoring instead of token generation, kept fully local. Source: CompleteAITraining.

The single most relevant item in today's brief to your actual stack. Your llama-server -hf ggml-org/Kev-4B-GGUF serving /v1/systemone on :8080 is exactly the shape this endpoint wants — first in line to test which models grow System One heads, and whether single-pass scoring beats your current setup on latency.

Decision 3.0: six open multimodal decision models land on Hugging Face

Six Apache-2.0 checkpoints — d3, d3-flash, d3-mini, d3-nano, d3-lite, d3-edge — appeared under the vllm-sr org on Oct 10, built on Qwen3.5 and Qwen3.8 backbones. They answer choice/yes-no/score questions with per-option probabilities instead of generated tokens; five of six read images and video. This is the vLLM project's open decision layer for its Semantic Router — the closest public cousin to a Jev-style decision head in GGUF-friendly territory. Source: OrcaRouter.

Open decision heads with per-option probabilities are calibration gold — exactly the shape your frozen question types want. d3-nano/d3-edge are the ones to eyeball for the talker lane; if any of these quantize cleanly to GGUF, they go straight into the Glimmer-vs-Qwen eval conversation.

Bloomberg ships enterprise MCP; four MCP servers land in 24h

Bloomberg released a Bloomberg Enterprise Model Context Protocol for its Data Licence Plus service — 100M+ securities and 50,000+ fields plus semantic search and reusable "Skills" (outlier detection, sanctions screening) exposed to client AI agents through a standard interface. One of four MCP server launches in the same window (ObservePoint, Waters AI Link, Fastweigh also shipped). Source: ITBrief.

Every data vendor is becoming an MCP server — the enterprise land grab is the MCP-everywhere momentum story of the day. Your move is the inverse: MCP servers for your own box, so Hermes queries your stack the way Bloomberg's clients query theirs.

New-model snapshot — access & licensing at a glance

Click a column header to sort. The licensing column is the story: open vs closed is now the axis that matters.

Model / Service	Vendor	Access	License	Price
Muse Wonder	Meta	Model API preview	Closed	$4 in / $20 out per 1M tokens
Muse Wonder Full	Meta	Muse Code subscribers only	Closed	$50 or $200/mo tiers
Decision 3.0 (d3 family)	vllm-sr / vLLM	Hugging Face	Apache-2.0	Free / open weights
System One endpoint	llama.cpp	Local	MIT (llama.cpp)	Free

BEAT 02

Hardware

4 stories

NVIDIA reportedly ends RTX 5090 shipments, redirects GB202 to RTX Pro

Leakers MEGAsizeGPU and hongxing2020 report NVIDIA has stopped allocating GB202 dies to GeForce RTX 5090 production, sending all future supply to RTX Pro workstation/server cards; NVIDIA has not confirmed, and no end-of-life date is announced — treat as highly likely but unconfirmed. The numbers that matter for local inference: the 5090 is the only consumer NVIDIA card with 32GB VRAM; MSRP was $1,999, but the cheapest new listings now run $7,178–$7,443 (a 259–272% markup, ~$233/GB of VRAM). Rumors point to a 24GB RTX 5080 refresh filling the gap. Sources: Overclock3D, GameGPU.

The worst possible card to lose for local inference — the only 32GB consumer option — and street prices are already at 2.6× MSRP. Run the calculator below and weep; then remember your dual-A770 pool cost a fraction of one scalped 5090 and start appreciating it more.

OneXPlayer Super X Pro: 192GB unified memory, "up to 300B" local inference

Previewed today (Oct 13): a 2-in-1 built around the 16-core/32-thread AMD Ryzen AI Max+ PRO 495 (5.2GHz boost, Radeon 8065S iGPU with 40 RDNA 3.5 CUs, 55 TOPS NPU) with up to 192GB of LPDDR5X-8533 at 273GB/s — up to 160GB dynamically allocatable as VRAM. The company claims local inference for models up to 300B parameters, making it the most VRAM-rich consumer-class local-inference box previewed this year. Caveat: vendor claim, not a tested benchmark. Source: Techaeris.

Strix Halo-class boxes keep proving the thesis: the future of big local models is unified memory, not discrete VRAM. 160GB of addressable VRAM in a 2-in-1 would swallow your entire dual-GPU VRAM math whole — vendor claim for now, but the direction is exactly right.

Intel names "D-Series" (Crescent Island): air-cooled AI inference at 350W

Intel gave its Crescent Island datacenter AI chip a product name: 1.3 petaflops FP8, 1.5 TB/s memory bandwidth, 350W, standard PCIe form factor with no liquid cooling required, partner cards carrying up to 480GB of LPDDR5X. For reference, AMD's comparable MI350P PCIe card peaks at 2.3 PFLOPS FP8 but draws 600W — Intel is pitching efficiency and drop-in datacenter retrofit instead of raw peak. Source: AIStockWire.

Datacenter-only, so no home-lab shopping — but the framing is the tell: Intel conceding peak FLOPS and selling efficiency plus air-cooled retrofit instead. The inference market is segmenting into "raw peak" vs "watts and thermals," and Intel just picked its lane.

AMD announces Helion: an agentic-AI rack with up to 32,768 CPU cores

AMD introduced Helion, an Open Compute Project rack blueprint purpose-built for running AI agents: a CPU-only config with up to 32,768 sixth-gen EPYC (Zen 6) cores, plus an enterprise config pairing EPYC blades with Instinct MI350P GPUs for combined agent execution and local inference. Same-day context: Microsoft said it will make its own Maia chips its main AI inference system, with its cloud chief noting long-running agents are driving "a huge rise in terms of the need for CPUs." Partner Helion systems target November 2026 (Cisco, Dell, HPE, Lenovo, Supermicro signed on). Source: AIStockWire.

The agentic era's bottleneck isn't just GPUs anymore — it's cores to run the loops. Microsoft admitting long-running agents are driving CPU demand is the industry quietly validating what your Hermes setup already knows: agents are a CPU workload wearing a GPU costume. Your 9800X3D approves.

GPU VRAM value calculator — $ per GB

The 5090 story in one number. Drag the price, pick the VRAM, and see what a gigabyte actually costs you. Preloaded with the reported street price.

Street price: $7310 VRAM: GB

~$228.44 / GB

5090 @ MSRP $1,999 → $62.47/GB5090 street $7,178–$7,443 → ~$233/GB

At 272% markup, each gigabyte of 5090 VRAM costs nearly 4× its MSRP rate. Unified-memory boxes (see the OneXPlayer above) are eating this market's lunch.

BEAT 03

Software

4 stories

Bazzite ships 44.20261012, then a .1 follow-up a day later

The Fedora-atomic gaming distro rolled kernel 7.2.9-ogc3, the Nvidia open driver 615.78.08, and InputPlumber 0.82.0 with a long list of controller fixes (DualSense, Legion Go 2/S, Steam Deck, Horipad). The Oct 13 point release adds more kernel tweaks. Sources: GamingOnLinux, Linux Compatible.

You don't run Bazzite, but the InputPlumber 0.82.0 fixes and the open-driver bump are the parts worth stealing. Atomic gaming distros are now shipping fixes faster than most of us update our dotfiles.

Strata v0.1.43.1: up to 38% faster local inference on a 125B sparse MoE

The MIT-licensed single-model engine (~22.6k stars) fixed the 125B speed regressions and crashes from 0.1.43. Claimed gains: +38% decode on a 16GB Radeon (AI PRO R9700, IQ3_XXS), and on Intel's Arc Pro B70 a 4K prompt dropped from 8.5s to 3.3s — byte-identical outputs to 0.1.42. The engine targets one 125B MoE that activates ~6B params/token; independent write-ups clock it ~2–2.6× faster than llama.cpp on the same hardware. Source: Linux Compatible.

This one is aimed squarely at your lane: an MIT engine that makes a 125B model behave like a 6B one at 2×+ llama.cpp's speed, with real Arc benchmark numbers. If the dual-A770 rig ever hosts a big MoE, Strata is the first thing to bench.

Linux 7.3-rc7: mount hardening, ext4 deprecation work, AMD errata workaround

The release candidate is RC-only for now — a watch item for Arch rolling, no action until stable. The stables you can actually run: 7.2.10 / 6.18.56 landed Oct 11 with KASLR and TCP use-after-free fixes; XanMod and Liquorix have already re-rolled (7.2-18 with zswap/zram tweaks). Source: Linux Compatible.

RC means "read the changelog, touch nothing." The interesting meat is mount hardening and the slow ext4 funeral — the stables are the boring-but-good security fixes.

Python 3.15 stable: lazy imports, UTF-8 by default, 7–8% faster JIT

Just outside the 24h window (Oct 9) but too relevant to skip: PEP 686 makes UTF-8 mode default, PEP 810 adds explicit lazy import, plus new frozendict and sentinel builtins, and an experimental JIT claiming 7–8% over the standard interpreter on x86-64 Linux. Source: Linuxiac.

lazy import is the feature every slow-starting Python tool has been begging for — including half the scripts in a home-lab stack. The JIT gains are modest, the import story is the headline.

Strata vs llama.cpp — decode speed on identical hardware

Independent write-ups clock Strata at ~2–2.6× llama.cpp's decode for the 125B MoE; hover the bars, or flip the engine to compare. Values normalized to llama.cpp = 1.0×.

llama.cpp (baseline)

1.0×

Strata 0.1.43.1

2.3×

~2.3×

Arc Pro B70 4K prompt

8.5s

Arc Pro B70 4K prompt

3.3s

Read it as: same model, same VRAM, ~61% less waiting on a 4K prompt. The +38% decode figure is on a 16GB Radeon at IQ3_XXS.

BEAT 04

Games

4 stories

Battlefield 6 Title Update 2.1.1.0: the polish pass

Almost entirely bug fixes: spectator-mode camera transitions and scoreboard/End-of-Round viewing repaired, store tiles rendering correctly, and a long tail of PC-specific fixes (Esc now closes the Quit prompt properly, correct mouse prompts when panning, attachment-alignment corrections). Zero balance changes. Source: DSOGaming.

The most boring-good kind of patch: no meta shakeups, just the UI behaving. Install, forget, keep shooting.

Hell is Us 1.8 adds 10 new Hollow Walker variants

The headline of this week's Game Pass update roundup: 10 new enemy variants aimed squarely at the repetition that sets in during the game's later hours. Also in the roundup: Minecraft Dungeons II fixed blurry visuals on high settings plus crashes, and Beast of Reincarnation added zone-skipping to speed up New Game+ runs. Source: TrueAchievements.

Directly attacks the game's biggest complaint — mid-game enemy déjà vu. If you bounced off the repetition, this is the patch to come back for.

Warhammer Survivors: Steam demo live, Nov 10 launch, day-one Game Pass

A Vampire Survivors-style bullet-heaven from Auroch Digital, Poncle, and Games Workshop spanning both 40K and Age of Sigmar. The demo is playable now ahead of the Nov 10 launch (Steam, GOG, PS5, Xbox Series X|S, Switch/Switch 2, mobile) with day-one Xbox Game Pass; promises 75+ weapons/relics/Arcana cards and 40+ unlockable characters. Source: Game8.

Poncle co-developing a 40K bullet-heaven is the most "why didn't this exist sooner" collab of the year. The demo is free, the genre eats evenings, and it runs on a toaster — the Ark's 55 inches will do nicely.

2XKO's final content update lands Oct 27 — Sett joins Samira

Riot's tag fighter gets its last content drop Oct 27 with Sett joining Samira; active development ends in December, though servers stay online beyond 2026. Source: Game8.

A graceful-ish sunset: one more character, then maintenance mode. Worth a final lab session before the roster freezes.

On the radar — dates worth marking

Click a column header to sort.

Date ▲	Title	What
Oct 27	2XKO	Final content update — Sett joins; dev ends in Dec
Nov 10	Warhammer Survivors	Full launch (Steam/GOG/PS5/Xbox/Switch/mobile) + day-one Game Pass
Now	Warhammer Survivors demo	Playable on Steam today
Oct 12	Battlefield 6 TU 2.1.1.0	Polish patch — UI, spectator, PC input fixes
Oct 12	Hell is Us 1.8	10 new Hollow Walker variants

BEAT 05

GitHub Repos

4 repos

morluto/rea — reverse-engineer anything with agents

A single MCP server + CLI (npx rea-agents setup) that hands a coding agent (Claude Code, Codex, Cursor, Copilot, Gemini CLI, OpenCode…) decompilation and tracing tools for native binaries, Electron apps, .NET assemblies, Android APKs, firmware, websites, and HAR captures. The agent decompiles, traces a feature, returns evidence with caveats — fully local, the binary never leaves your machine. #1 on GitHub Trending today; ~188.8k stars, roughly 167k of them in the last 30 days — one of the fastest-growing repos GitHub has ever seen. REA 4.1 (Oct 6) added APK analysis via JADX, firmware via Binwalk/Unblob, and IDA Pro as a backend. MIT, TypeScript.

Local-first, MCP-native, built for exactly your agent-harness world — and growing faster than almost anything in GitHub history. The 4.1 APK/firmware additions turn it from a demo into a toolbox. This is the repo to clone this week.

mksglu/context-mode — stop drowning your agent's context window

A local MCP server that intercepts tool output before it floods an agent's context — sandboxed execution returns lightweight pointers instead of raw payloads (claims up to 98% context reduction), persists session memory in SQLite FTS5, and installs lifecycle hooks plus a routing skill across 17 platforms (Claude Code, Codex, Copilot, Cursor, OpenClaw, Zed…). #3 on GitHub Trending today; ~27.7k stars and climbing steadily. TypeScript.

Context amnesia is the #1 killer of long agent sessions, and this attacks it at the plumbing level. Speaks directly to your trajectory-logging and calibration work — the underconfidence canary is, at heart, a context-cost measurement.

lyogavin/airllm — 2.8T-parameter inference on a 4GB GPU

A layer-sharded inference library that streams model weights one layer at a time from disk — absurd-scale models on tiny GPUs with no quantization: 70B on 4GB VRAM, Kimi K3 (2.8T) in 3.72GB, DeepSeek-V3 (671B) in ~12GB, Qwen3.8-Flash-Next (125B) in 5.95GB. A September update added training huge models on small VRAM (125B under 6GB). #7 on GitHub Trending today; ~38k stars. Python, Apache-2.0, one-line pip install.

This is the VRAM math you live for — and the layer-split philosophy you already endorsed over tensor split, running in production. The Kimi K3-on-3.72GB trick reframes your "biggest model fitting ~30GB" directive entirely. Keep one eye on this during the Glimmer 30B eval.

microsoft/BitNet — official 1-bit LLM inference

Microsoft's official framework for 1.58-bit (ternary) LLMs: hand-tuned kernels delivering 2.4–6.2× speedups and up to 82% energy savings on x86 CPUs — a 100B BitNet model runs on a single CPU at human-reading speed. Also ships the official BitNet-b1.58-2B-4T model, 1-bit embedding models (0.6B, 270M), and GPU kernels. #8 on GitHub Trending today with a steady 72h resurgence; ~41.1k stars. C++/Python, MIT.

The most radical answer to your VRAM ceiling: if models go 1-bit natively, the 29.4GB math stops being the binding constraint. A 100B model at reading speed on bare CPU is the kind of result that makes GPU shopping look quaint. Watch it for the d1-3B test's sake.

Trending board — today's ranks, sorted by stars

Click a column header to sort. Star counts read live from GitHub this morning — they move by the minute.

Repo	Lang	Stars ▲	Trending	Why it fits
morluto/rea	TypeScript	188.8k	#1	Agentic reverse engineering, fully local, MCP-native
microsoft/BitNet	C++/Python	41.1k	#8	1-bit inference — the VRAM-ceiling killer
lyogavin/airllm	Python	38.0k	#7	2.8T on 4GB — layer-split VRAM math
mksglu/context-mode	TypeScript	27.7k	#3	98% context reduction for agent sessions

♥ The Wedding Countdown

410 days to go — Saturday, November 27, 2027

the lodge, up in the hills — the whole weekend is yours, Friday the 26th through Sunday the 28th. Four hundred ten days out, mid-October light doing its golden thing: this is the season for long walks and slow decisions.

On record: venue weekend package booked (Fri Nov 26 – Sun Nov 28, 2027) · food-cart crown still held by Rip City Grill at $14/person · vendor quotes not yet requested.