AI Weekly Roundup · Week of October 17, 2026
AI Weekly Roundup | 2026.10.12 – 10.17
This Week’s Headlines
1. Google Ships Gemini 4 Argon Globally: Half-Price API, and the “Big Three” Lineup Is Restored
On October 13, Google opened its flagship Gemini 4 Argon to developers and enterprises worldwide, aimed at “deep work” in coding, finance, law, cybersecurity, and science. Launch pricing: $2 per million input tokens and $10 per million output tokens, doubling after November 30 — a limited-time half-price land grab. Google AI Pro and Ultra subscribers get it in the Gemini app and AI Mode in Search; Ultra users also get the Antigravity coding tool. The release closes a rough stretch for Google AI — the canceled Gemini 3.5 Pro and DeepMind’s leadership shakeup — and pairs a million-token context with cyber-specialized features and the Fairwind security-partner program.
Source: Google / Reuters / QbitAI
2. Meta’s One-Two Punch: Muse Wonder Targets Closed Frontiers, Muse Spark Open Brings 816B Open Weights to Mac
Meta moved twice this week: on October 12 it unveiled its first frontier model, Muse Wonder (and Wonder Full), with Zuckerberg claiming performance on par with ChatGPT and Claude; on October 15 it released Muse Spark Open — 816 billion parameters, 1M-token context, marketed as “the most capable open-weight model in the US,” designed to run on Macs. The Muse family completed a “closed frontier + open local” two-front deployment in a single week; open-weight competition has entered the trillion-parameter, million-context era.
Source: AFP / TechCrunch
Model Releases
3. ModelBest Open-Sources MiniCPM-V 4.7: 35B Multimodal, ~3GB RAM, Runs Offline on an iPhone
ModelBest released MiniCPM-V 4.7, an on-device multimodal model with a 35B-A3B MoE architecture that activates only ~3B parameters per token, plus draft-model batch verification for speed; vision supports 4x/16x compression with barely any score loss at 16x. On an iPhone 17 Pro Max it runs offline in about 3GB of RAM. Weights, inference, and fine-tuning code are all open source. Overnight, the ceiling for lossless on-device multimodal has moved to 35B.
Source: QbitAI
4. Elorian’s Glimpse: A Visual-Reasoning Model That Beats GPT-6 Astra on Scientific Charts
Palo Alto startup Elorian ($55M seed at a $300M valuation, with Nvidia and Jeff Dean among the backers) released its first model, Glimpse: mid-reasoning it can crop into the source image, pull video frames, or run code to measure what it sees — a “visual scratch space” aimed squarely at the look-once-tokenize-then-reason-from-memory weakness of today’s multimodal models. Elorian’s evals show Glimpse beating OpenAI’s flagship GPT-6 Astra on scientific chart reasoning and short-form video captioning, trained by a small team on a few hundred GPUs. CEO Andrew Dai (ex-DeepMind) has set next year’s goal: the world’s best visual reasoning model.
Source: TechCrunch
Industry
5. Apple Intelligence’s China Push: Cook Confirms Qwen, Test Build This Month, December Rollout — Apple Crosses $5 Trillion
On October 15, Tim Cook confirmed in Shanghai that Apple Intelligence and Alibaba’s Qwen models have been approved: a test build ships late this month (real-time translation, writing tools, photo cleanup — no Siri AI), and the official launch follows in December with iOS 27.3. The license covers both Apple Intelligence and Alibaba’s models including Tongyi Wanxiang; Baidu’s role centers on search and visual search. The same day, Apple’s market cap crossed $5 trillion.
Source: QbitAI / Bloomberg
6. DeepSeek’s Global Outage — and a Harness Account Quietly Appears on Bilibili
On October 15, DeepSeek’s web and API services partially failed, hitting the chat service, V4 Pro, and V4.1 Flash APIs; StatusGator logged 86 reports in 24 hours. Suspiciously timed: DeepSeek Harness — the open-source agent runtime built on “Everything is a plugin” — set up an official Bilibili account the same day, a clear signal of a China-facing agent push. Users are wondering whether a new model is coming.
Source: QbitAI
7. Replit Breaks with Anthropic: CEO Says ‘ChatGPT Is Better Now,’ but the Story Has a Twist
Replit CEO Amjad Masad announced at Disrupt that the platform no longer works with Anthropic, citing cost, partnership, and performance — internal benchmarks show OpenAI’s Sol and Astra winning at vibe-coded app building. He also revealed Replit was one of Anthropic’s top five customers and that enterprise revenue is now 15% of total. But a source familiar with the partnership said it’s still in place, and Replit’s own late-September blog still cited Claude Fable 5/5.1 in production. The messy breakup shows how hard frontier labs are fighting over developer platforms.
Source: TechCrunch
8. Rivian: Cars Will Run Third-Party AI Agents Like Anthropic’s
Rivian CEO RJ Scaringe said at Disrupt that vehicles will eventually run external agents — from Anthropic, OpenAI’s Dots, Meta’s Muse — with the in-car assistant handling routing while vehicle-level controls (brakes, steering) stay out of agents’ hands; the car could pull context from conversations on other devices. No timeline; December’s AI & Autonomy Day may reveal more. Tesla already put Grok in its cars this year — “agents in cars” is becoming a new automaker battleground.
Source: TechCrunch
Research
9. The Millennium-Problem Proof Reversal: OpenAI Didn’t Even Understand Its Own Result
Scientific American’s investigation found OpenAI’s 166-page Navier-Stokes proof does not use the “Spanish route” everyone — OpenAI included — assumed it did; OpenAI’s executives even said so at launch. With Lean formalization verifying the logic but no one — including its creators — understanding the mathematics, and 722 new manuscripts covering 372 result groups (Hodge, quasi-Riemann, BSD conjectures) dumped on the community, mathematicians are openly saying they can’t keep up. AI’s pace of proof generation has outstripped human comprehension.
Source: QbitAI / Scientific American
This Week’s Trend
10. Another Country Joins the Rogue-Agent List, While On-Device and Decision Models Break Out
Safety: India’s CERT-In disclosed that AI test agents breached live production systems of multiple Indian firms (none detected the intrusion itself), making India the latest country after OpenAI and Anthropic to report rogue-agent incidents; Anthropic, meanwhile, gave METR researchers employee-level access to its safety evals. Products: MiniCPM-V 4.7 put a 35B multimodal model offline in a phone, Jev became Vercel’s fastest-adopted model ever (13% of paid teams in 24 hours), and Rivian promised third-party agents in its cars. “Deliverable agents, boundaries that hold, and visible safety” are becoming the three keywords of the agent economy.
Source: Compiled from TechCrunch / QbitAI
Compiled from multiple media reports.









