Morning Tech Brief — October 14, 2026

BEAT 01

AI Tools

4 stories

Google launches Gemini 4 Argon — flagship model at a fifth of the flagship price

Google released Gemini 4 Argon on Oct 13 after months of delays and a DeepMind leadership shakeup, pitched at "deep work" on complex multi-step tasks (coding, finance, law, cybersecurity, science). API pricing opens at $2/$10 per million input/output tokens — roughly a fifth of GPT-6 Astra and Claude Opus 5.5 — but the intro rates double after November 30. Early reception is unusually warm: HN commenters called it a "clear step change" for reasoning, and Google claims internal teams raised agentic code submissions 35% while cutting rollbacks by a third. Caveat: image-to-HTML tests put it a tier below Opus 5.5 and GPT-6.1 Sol. Also: Google is offering 10,000 scientists six months free on the Antigravity coding tier. Source: Reuters.

The most relevant cloud-model news in months for your stack. At a reported $0.41/task on arena tests vs $3.51 for competitors, Argon resets the cost bar any paid fallback gets weighed against — your Qwen3.6-27B talker lane and the Nous free pool now have a cheap cloud neighbor. Watch whether the sub-Nov-30 rates survive the honeymoon.

Nous Research raises $90M Series B for "Hermes for Businesses"

Nous Research raised $90M (a $1.5B valuation) to build out "Hermes for Businesses" — letting companies choose models by price/goal, run privately, and manage company knowledge, with SSO and audit logs planned. The core Hermes Agent stays open-source under MIT. Nous claims 24M+ clones, 22M+ downloads since February, and ~2.5% of global AI token usage (company estimates, not independently audited). Source: TechAIWire.

Your entire cloud tier runs on Nous's free models, so this raise is the single most important signal for whether that free pool survives, improves, or gets paywalled. The "choose models by price" framing suggests the free tier persists as the funnel — good news for your setup. And if the enterprise pivot starts eating Nous's attention, your Hermes harness is MIT anyway — you keep the code regardless.

Muse Wonder follow-up: Artificial Analysis scores it, Muse Spark open release still missing

Fresh analysis of Meta's Oct 12 pair (covered yesterday): Artificial Analysis scores Muse Wonder Full at 56 — tied with Claude Sonnet 5.5, #2 behind Opus 5.5's 58 — with Meta claiming a big lead on long-horizon coding (SWE-Marathon 57.5 vs Opus 5.5's 46.3). Only the weaker Wonder is API-accessible ($4/$20 per 1M, 1M context, six reasoning-effort settings); Wonder Full is Muse Code-subscription-only ($50–$200/mo). And the open-weights Muse Spark release Meta promised in September still hasn't shipped. Source: VentureBeat.

The missing Spark release is now the loudest data point: Meta isn't bringing open weights to the frontier, full stop. For work-order #4, Wonder's numbers are the bar your Glimmer 30B A/B has to clear — and the licensing column keeps doing Glimmer's marketing for it.

Alibaba opens sign-ups for Qwen's in-app Personal Agent (China)

Alibaba opened invite-batch sign-ups on Oct 13 for "Personal Agent by Qwen" — an agent embedded in the Qwen App (China only) that pursues refunds, files flight-delay compensation claims, books services, and handles car maintenance, adapting to instructions and account info while users watch, intervene, and stop tasks. No general availability date, pricing, or success-rate data disclosed. Source: RuntimeWire.

The first real-world agentic tool-use deployment from the Qwen house — the model family anchoring your entire local stack. Refund-chasing and claim-filing is the unglamorous stuff agents actually need to be reliable at. Meanwhile Qwen4 sits in pre-training with no release date ("coming soon" since Sept 22) — watch this space, not your breath.

Fallback-cost calculator — what your cloud neighbor costs per month

Argon's intro rate is the story: drag daily token usage and see the monthly delta against Wonder and the ~$10/$50 flagship class. Rates: Argon $2/$10 (intro, doubles after Nov 30), Wonder $4/$20, flagship-class ~$10/$50.

Daily input: 5M tokens Daily output: 1M tokens

Gemini 4 Argon (intro)

Muse Wonder

Flagship class

30-day month. Local models (Qwen3.6-27B, Nous free pool) cost $0 — the calculator exists to price the "oh no, the free tier died" scenario.

Frontier access board — click a column header to sort

The licensing column is still the story: open vs closed is the axis that matters.

Model / Service	Vendor	Access	License	Price
Gemini 4 Argon	Google	API	Closed	$2 in / $10 out per 1M (intro)
Muse Wonder	Meta	API	Closed	$4 in / $20 out per 1M
Muse Wonder Full	Meta	Muse Code sub only	Closed	$50 or $200/mo tiers
Hermes Agent	Nous Research	Open source	MIT	Free / self-host
Qwen Personal Agent	Alibaba	In-app, invite (China)	Closed	Undisclosed

BEAT 02

Hardware

4 stories

AMD reveals EPYC "Verano": 192 Zen 6 cores, first-ever LPDDR5X on EPYC

Shown at the OCP Global Summit on Oct 12: Verano is a Zen 6 EPYC 9006-family chip purpose-built for agentic AI — six 32-core Zen 6C CCDs (192 cores), and the first EPYC with an LPDDR interface via the new SOCAMM2 form factor. It connects to Instinct MI500 GPUs over up to 112 Gbps xGMI links on the SP8 platform, shipping in 2027. Source: Wccftech.

The datacenter is converging on the conclusion you already live by: memory bandwidth and capacity are the real AI bottleneck, not raw compute. LPDDR5X on a server EPYC is the enterprise echo of the unified-memory logic behind Strix Halo — and of the VRAM math you run on the dual A770s. Pairs nicely with yesterday's Helion story: AMD is all-in on agents-are-CPU-workloads.

AMD launches Ryzen AI Halo Developer Platform: 192GB unified, 160GB allocatable as VRAM

Built around the Ryzen AI Max+ PRO 495 (16 Zen 5 cores/32 threads, 5.2 GHz boost, 55W default TDP, Radeon 8065S with 40 RDNA 3.5 CUs, XDNA 2 NPU at 55 TOPS). Up to 192GB of 256-bit LPDDR5x-8533 (273 GB/s), with up to 160GB allocatable to the GPU as VRAM — plus PCIe 4.0, dual 40 Gbps USB4, 2TB SSD, 10 GbE, Wi-Fi 7. Source: Blockchain.news.

The strongest local-inference consumer platform of 2026, full stop. 160GB of addressable GPU memory on one box would swallow models your dual 16GB A770s can't touch — including the Muse Glimmer 30B evaluation you have parked as work-order #4. A concrete data point for the "bigger memory pool vs. dual discrete cards" math you keep running.

NVIDIA RTX Spark AI PCs land in October: Lenovo and Acer first, from $2,599

The Arm-based RTX Spark platform (20-core Grace CPU + Blackwell RTX GPU, up to 128GB unified memory) ships this month — Lenovo and Acer confirmed October launches, with Asus, Dell, HP, MSI, and Microsoft Surface also on board. Starting price in the US: $2,599. Related: Mechrevo's Master Road X10 mini PC (also RTX Spark-based, covered Oct 12) adds a proprietary "AI SSD" that offloads model layers from DRAM to NAND — Mechrevo claims it runs 80B–122B-parameter models and boosts supported model size up to 5.7×. Sources: Tech-Insider, Notebookcheck.

NVIDIA's answer to the same capacity problem you fight on a 32GB-combined A770 pair. The Mechrevo AI-SSD trick is the one to watch — if SSD layer-offload works without murdering tokens/sec, it rewrites the VRAM math on every rig like yours. Treat the 5.7× claim as vendor marketing until someone independent benches it.

RTX 6080/6090 rumor: GDDR7 shortage may force NVIDIA to cut VRAM configs

Per an Oct 7 TLDVTech report cited by Tech-Insider: the RTX 6080 may ship with 16GB GDDR7 (down from a rumored 24GB) while the RTX 6090 drops from a rumored 48GB to 32GB at a price approaching $3,000. The driver is a GDDR7 crunch: 2GB modules reportedly hit ~$15 apiece in some batches, Micron discontinued its 2GB GDDR7 modules, and Micron's CEO says memory capacity is ~75% sold out heading into 2027. Unconfirmed — treat as rumor-grade. Source: Tech-Insider.

If true, flagship VRAM is getting smaller — the exact opposite of what local inference needs. It reframes your A770 pair's 32GB combined (and the Arc Pro B70's 32GB at $949) as surprisingly competitive per-gigabyte: a $3,000 card with 32GB makes memory-poor NVIDIA flagships a tax on patience. The unified-memory boxes above are eating this market's lunch, and NVIDIA knows it.

Addressable memory pools — the only spec that matters for local inference

One bar per platform. The rumor column is the trend: NVIDIA's next flagships may match your A770 pair's combined VRAM — at roughly 5× the money.

Dual Arc A770 (your rig)

32 GB

Arc Pro B70 (32GB, $949)

32 GB

RTX 6090 (rumored)

32 GB*

~$3k*

RTX Spark AI PC (128GB)

128 GB

$2,599

Ryzen AI Halo (160GB alloc.)

160 GB

—

* Rumor-grade: RTX 6080 possibly 16GB, 6090 32GB. GDDR7 at ~$15 per 2GB module is why.

BEAT 03

Software

4 stories

OpenCode 2.0: terminal, desktop, and web on one background service

OpenCode's major release (from Anomaly, three months after public beta) moves every client — terminal UI, desktop app, web interface, and scripts on the new @opencode/client package — onto one shared background service that starts automatically and owns sessions, config, integrations, permissions, and tool execution. opencode pair lets other hosts reach the same service. Also new: a rebuilt API, an embeddable SDK (@opencode/sdk), and Code Mode, which lets the agent write JavaScript to chain several MCP tools; subagents and shell commands can run in the background. Breaking change: 1.x plugins don't work in 2.0 and need rewriting. Source: AlexTech.ai.

The open-source coding agent everyone watched in beta just went architecture-first: one stateful service, many thin clients — which is also the shape you'd want for a phone-accessible Hermes console. Code Mode's "agent writes JS to chain MCP tools" is the closest thing to your own directive-passing patterns. Plugin rewrites are the tax; nothing good is ever free.

VS Code 1.141.1: critical 9.6 RCE — patch now

Microsoft's Oct 14 security update closes a critical remote-code-execution hole rated 9.6. That's about as loud as CVE severities get before the perfect 10. Source: Linux Compatible.

If VS Code is open on the home PC right now, update it before coffee. The 9.6 RCE is the kind of thing that turns an editor into a foothold — and the patch board below is your triage list for the rest of the week.

Strata 0.1.44.1: Mac support, Unsloth quants, NVFP4

The MIT-licensed single-model engine (~22.6k stars, covered yesterday for its +38% decode claim over llama.cpp) keeps shipping: 0.1.44.1 adds Mac support, Unsloth-format quants, and NVFP4 support — two days after the 0.1.43.1 hotfix that fixed 125B speed regressions. Source: Linux Compatible.

Strata's cadence is the story: hotfix Monday, features Wednesday. NVFP4 support matters because it unlocks Blackwell-class inference paths; Unsloth quants matter because that community ships some of the cleanest low-bit weights around. Still the first engine to bench against llama.cpp when a big MoE lands on the A770s.

Linus Torvalds used AI to untangle obscure kernel code — and found bugs from 2013

Faced with the dense, multi-level assoc_array code and a proposed fix he couldn't judge, Torvalds turned to LLM assistance: the models helped decompose monolithic routines into traceable functions, then a userspace test harness (built to verify without reboot cycles) confirmed the original bug and surfaced three more defects — all dating to the very first assoc_array commit in 2013. The patches live in a topic branch in his personal tree, likely headed for the 7.4 merge window. Sources: WebProNews, Linux Compatible.

The man who called AI "90% marketing, 10% reality" just used an LLM as a refactoring partner on code he couldn't untangle himself — and it worked. The kicker is his method, not the models: decompose first, verify with a real harness, never trust the output. That's the same evidence-first protocol you run on Hermes, straight from the kernel's top.

Patch board — this week's triage, sorted by severity

Click a column header to sort. When in doubt, patch the editor first and the kernels second.

Severity ▲	Patch	Why now
CRITICAL — 9.6	VS Code 1.141.1	Remote code execution in your editor; update today
HIGH — kernel	Ubuntu 24.04 USN-8917/8918	~1,050 CVEs each, incl. AMD SEV-SNP guest-memory hole (CVE-2023-20585)
HIGH	Chromium / Fedora-Debian	258 CVEs across the Linux security roundup
MEDIUM	Samba 4.24.8	Ten bug fixes incl. SMB3 signing patch — schedule it
LOW — routine	Arch: linux 7.2.10.arch1-1	In core-testing now; KASLR + TCP use-after-free fixes ride along

BEAT 04

Games

4 stories

Steam breaks 43.5M concurrent users — third record of 2026

SteamDB logged 43,505,724 concurrent users on Sunday Oct 11 — the platform's third record this year (after ~42M in January and ~42.5M in March). About 14.46 million were actively in-game at the peak, also a record; the rest were logged in, chatting, browsing, or shopping. No single title drove it: Deadlock peaked at 262,145 after its Sept 30 update, Aion 2 hit ~397,905 at launch, with CS2 (1.4M), Dota 2 (1.03M), and PUBG (713k) doing the steady lifting. Valve hasn't commented. Source: Polygon.

PC gaming's platform-of-record just lapped its own record for the third time this year, and the peak nearly doubled over the last six. On your 55 inches of Ark, that means fuller lobbies and more devs prioritizing the PC build. Worth noting: 29 million of those "users" had Steam open and weren't playing — chat-and-shop counts too.

Battlefield 6 Season 5 "Siege of Vegas" is live — Vegas Strip, Area 51, four weapons

Correction to yesterday's framing: 2.1.1.0 isn't just a polish patch — it's the Season 5 opener. Two new maps (Vegas Strip: casinos, rooftop ziplines, storm drains; Area 51: jets, tanks, underground tunnels, a bring-downable control tower), four weapons (Mk 24 AR, RSPR PDW, Z84M belt-fed LMG, ELR-40 sniper), Rapid Royale and Ranked Battle Royale in REDSEC, plus a major mobile anti-air buff (faster projectiles, less dispersion) and a 30% nerf to light-aircraft rocket splash. Free trial on PS5 and Steam Oct 15–19; the BF6 x Dead Space: Convergence event (16-player PvE Marker Incursion) runs Oct 20–Nov 7. Sources: DSOGaming, HowToVideoGame.

Yesterday's brief sold this short as a bug-fix pass — it's actually the biggest BF6 content drop of the year. The free trial starting tomorrow is the perfect excuse to drag friends into Rapid Royale on the Ark. And yes, Dead Space in Battlefield is a crossover nobody asked for and everybody will play.

Street Fighter 6 adds Arjun — 6'9" of Yoga Breathing and a Bollywood level-3

Capcom added Arjun on Oct 13 as the second Year 4 fighter (requires the Year 4 Character Pass): a 6'9" former Indian police officer framed for his partner's murder, fighting with "Yoga Breathing" — slow, powerful strikes that deal recoverable damage even on block, plus air projectiles and traps. C. Viper's Outfit 4 dropped alongside; Tifa (FFVII) is confirmed next, early 2027. Source: TwistedVoxel.

A brand-new archetype built around punishing block-happy opponents — and a level-3 super that reportedly ends in a Bollywood dance number. Lab him on PC if you own the pass; the real headline is Capcom keeping Year 4 this spicy.

over the hill launches today on Steam — SnowRunner with feelings

Funselektor (art of rally) released their cozy off-road adventure over the hill on Steam today, Oct 14: driving-not-racing in 1960s–80s-inspired 4x4s across Canada and Algeria, 1–4 player online co-op (no local), winches and planks for terrain traversal, deformable terrain, more regions promised post-launch. Consoles (PS5, Xbox, Switch 2) come in 2027. Source: Allkeyshop.

Today's freshest Steam release, and the pitch writes itself: the art of rally studio doing a 4-player co-op overland crawl. Made for a big screen and friends who don't take racing games seriously.

On the radar — dates worth marking

Click a column header to sort.

Date ▲	Title	What
Today	over the hill	Steam launch — cozy off-road co-op
Oct 15–19	Battlefield 6	Free trial (PS5 + Steam) — Season 5 is live
Oct 20–Nov 7	BF6 x Dead Space	Convergence event — Marker Incursion PvE mode
Nov 10	Warhammer Survivors	Full launch + day-one Game Pass (demo live now)
Early 2027	Street Fighter 6	Tifa joins the Year 4 roster
2027	over the hill	Console versions (PS5 / Xbox / Switch 2)

BEAT 05

GitHub Repos

5 repos

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: 315KB→5.4KB), persists session memory in SQLite with BM25 retrieval, and installs tool-routing hooks across 17 platforms (Claude Code, Codex, Cursor, Gemini, OpenCode, OpenClaw…). Trending #4 today (~28.7k stars); ELv2 license.

You literally installed this into Hermes's harness last night with all 11 tools wired — today's trending slot is confirmation of the pick. Now the real test begins: run ctx-stats through a real session and see whether your per-command savings match the 98% headline, or whether last night's 765-bytes-vs-1,800 is the honest number.

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

Pip package (from airllm import AutoModel) that runs giant LLMs on tiny GPUs by streaming model layers from disk — no quantization required: Kimi K3 (2.8T) in 3.72GB VRAM, DeepSeek-V3 (671B) in ~12GB, Qwen3-235B in ~3GB, Qwen3.8-Flash-Next (125B MoE) in 5.95GB. Trending #7 today (~38.6k stars); Apache-2.0.

The layer-split philosophy you endorsed over tensor split, running in production. This reframes your "biggest model fitting ~30GB" directive entirely — and it's worth a benchmark against llama.cpp on your own Qwen models, if only to know what the ceiling really is.

morluto/rea — reverse-engineer anything with agents

An MCP server + CLI that hands a coding agent decompilation and tracing tools for native binaries (Hopper/Ghidra/IDA), Electron apps, .NET assemblies, APKs, firmware, and PCAPs — returning findings with evidence and limits so the agent can explain and reimplement features without source. Analysis runs fully local; the binary never leaves your machine. #1 on GitHub Trending today (~228k stars, +32k today); MIT.

Back on trending's top step and still growing at absurd velocity. MCP-native, fully local, MIT — it reads like it was designed for your exact stack. The APK/firmware additions make it a genuine toolbox, not a demo. Clone it this week.

debpalash/VoiceStudio — fully-local ElevenLabs alternative

Voice cloning, voice design, video dubbing, dictation, transcription, and audiobook creation in 646 languages — all local, with an Electron desktop app plus a local API and MCP integrations so agents can drive it. Trending #3 today (~65k stars); AGPL-3.0 (fine for your never-distributed home use).

Local-first audio AI matches your north star exactly, and the MCP angle means Hermes could drive it. It's also a self-hosted alternative sitting next to MAI-Voice-2.1 for Tiny Hermes Buddy's voice side. AGPL keeps it home-only — which is where all your builds live anyway.

mattpocock/skills — Skills for Real Engineers

A curated pack of agent skills (/grill-me, /grill-with-docs, /tdd, /diagnosing-bugs, /code-review…) that target common agentic-coding failure modes: misalignment, verbosity, non-working code, entropy-heavy codebases. Installable across Claude Code, Codex, Copilot, Gemini CLI. #2 weekly trending (~290k stars); MIT.

Pure Hermes-harness fodder. The composable-skills pattern — shared glossary, red-green-refactor loops, anti-verbosity guardrails — is directly portable to how you skill up your own agents. Steal the pattern, skip the dependency.

Trending board — today's ranks, sorted by stars

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

Repo	Lang	Stars ▲	Trending	Why it fits
mattpocock/skills	—	~290k	#2 weekly	Composable agent skills — Hermes fodder
morluto/rea	TypeScript	~228k	#1	Agentic reverse engineering, fully local
debpalash/VoiceStudio	—	~65k	#3	Local voice cloning — Buddy's voice side
lyogavin/airllm	Python	~38.6k	#7	2.8T on 4GB — layer-split VRAM math
mksglu/context-mode	TypeScript	~28.7k	#4	You installed it last night — audit the savings

♥ The Wedding Countdown

409 days to go — Saturday, November 27, 2027

Four hundred nine days out, and the food-cart crown is still sitting with Rip City Grill at $14/person — weddings explicit, gratuity 18%. Late-October light, long walks, slow decisions: this is the season the planning book was made for.

On record: venue weekend package booked (Fri Nov 26 – Sun Nov 28, 2027) · food-cart #1 pick Rip City Grill $14/person · vendor quotes not yet requested.