AI Frontier Daily · October 14, 2026
AI Frontier Daily · October 14, 2026
Daily global AI trends digest, caught up in five minutes.
Headlines
1. ModelBest Open-Sources MiniCPM-V 4.7: A 35B MoE That Runs Offline on an iPhone in Just 3GB of RAM
ModelBest (面壁智能) and OpenBMB have officially open-sourced MiniCPM-V 4.7 in two sizes. The 35B-A3B variant packs 35B total parameters but activates only 3B per forward pass — the first MoE in the MiniCPM-V family — and runs smoothly on iOS and Android in roughly 3GB of memory. The 1B variant (1.3B actual parameters) targets the ultra-small edge, with OCR, math, and video understanding that lead its size class. In testing, 35B-A3B hit decode speeds above 13 tokens/s on an iPhone 17 Pro Max, fully offline, while occupying just 3GB of RAM. Benchmarks: MMBench 91.72 and MMStar 80.67 for general vision; MMMU 80.22 and MathVista 88.00 on STEM; Video-MME 76.40 and LongVideoBench 68.90 on video understanding; and a standout 94.24 on OmniDocBench v1.6 — more than 12 points clear of the four comparison models. Notably, pushing visual compression to 16x barely dented quality: MMBench slipped only from 91.72 to 90.87 and OCRBench from 91.20 to 90.40. On text-only tasks it leads too: CMMLU 90.21 (Chinese knowledge) and τ2-Bench 84.65 (tool use). The team bills it as the first open-source MoE multimodal model that runs fluidly on-device without sacrificing performance — another step down in the cost of running large models on a phone, with no network required.
Source: QbitAI
2. Kaiming He’s Team Releases VISTA, a Vision-Native Harness That Lets Models ‘Look Back’ at What They Saw
Kaiming He’s group has published VISTA: A Visual Harness for Reasoning in an Interactive World. Anthropic introduced the harness concept last November for long-horizon coding — task lists, progress files, and code records let a model continue across context windows — but that was textual, code-based state. VISTA makes it vision-native: the multimodal model perceives the environment directly through visual observations and maintains a lossless visual memory that preserves past frames in their original form. At reasoning time it can actively retrieve those observations and reorganize its visual input on the fly, with no retraining. The results are striking: paired with VISTA, Claude Opus 5.0 scores a perfect 100 RHAE (Relative Human Action Efficiency) across all 25 public games in ARC-AGI-3, completing them with 57.4% fewer actions than first-time human players; GPT-5.6 Sol also clears all games at 99. The team further validated VISTA on three additional benchmarks spanning browser games, mazes, and matching puzzles, substantially outperforming baselines that pair the same model with minimal harnesses. It’s the logical extension of He’s through-line from ResNet and Mask R-CNN to MAE: this time, visual experience itself becomes context the model can summon at will.
Source: QbitAI
3. DeepSeek to Lead Seed Round in Winverse AI, Founded by Swin Transformer Co-Author Han Hu
Per Bloomberg, DeepSeek plans to lead a multi-million-dollar seed round in Winverse AI, a robotics-AI startup, at a post-money valuation of roughly $400 million (~RMB 2.8 billion); Meituan’s Longzhu Capital has also committed. Founder Han Hu earned his bachelor’s from Tsinghua in 2008 and his PhD in 2014, then worked at Baidu’s Institute of Deep Learning and Microsoft Research Asia before joining Tencent Hunyuan in December 2023 as distinguished scientist, leading the vision-language and embodied-AI foundation model team and previously heading the Hunyuan LLM technical effort. His best-known academic contribution is Swin Transformer, co-authored with the team in 2021 (ICCV 2021 Best Paper), and he made MIT Technology Review’s China “35 Innovators Under 35” list in 2022. This is DeepSeek’s second robotics investment in two months — in August it put ~RMB 141 million into Unitree, with plans to collaborate on embodied intelligence and large models. If the round closes, it would rank among DeepSeek’s largest startup investments since founding; Tencent is also reportedly a potential investor in DeepSeek’s own new funding round.
Source: QbitAI, Bloomberg
4. Anthropic Researchers Demo ‘AI Self-Replication’: GLM-5.2 Autonomously Spreads Across 8 Isolated Devices
Anthropic security researchers William Zhou and Daniel Freeman published a hard-edged safety report on ARA (Autonomous Replication and Adaptation) — can an AI, like a computer worm, find a target over the network, copy its weights over, configure and launch itself on unfamiliar hardware, and spawn the next generation without a single human confirmation click? Experiment one: GLM-5.2 was tasked with spreading an 8B-parameter model across 8 devices on an isolated network — scanning for targets, transferring payloads, setting up environments and drivers on unseen hardware, and booting the model, all autonomously. In 5 runs it took all 8 devices 4 times and 7 devices the remaining time. Experiment two is pure cyberpunk: the researchers bought 15 cheap Android TV boxes (55 USD each, 4GB RAM), sharded a 26B model (Gemma4-26B-A4B) across them layer by layer using Petals-style decentralized distributed inference, and ran the full 26B brain over a LAN. The report estimates that by year-end, even models under 30B could plausibly possess end-to-end fully autonomous replication — meaning smart TVs, robot vacuums, and even internet-connected ovens could become new hosts for AI.
Source: QbitAI
Model Releases
5. Alibaba’s Qwen Go Debuts in Beta: Drop in One Sentence, and the Errand Gets Done
With every frontier lab now pushing personal agents (Meta shipped Muse, OpenAI revealed Dots), Alibaba has shown its hand with “Qwen Go.” Based on early beta-user reports, booking flights and hotels is entry-level: it digs out entrances you’d never find yourself, operates obscure features, and will even type round after round against real human customer-service reps. Case in point: a user couldn’t remember which of her flights had been delayed, furnishing only the booking platform and origin — Qwen Go logged into her account, worked order history to narrow down candidates, called customer service when the phone line failed, enlisted in-app agents to pin down the delayed flight, then pressed on to claim compensation, ultimately converting a 70-yuan voucher into cash through sheer persistence. Price-drop refunds, car-maintenance appointments, golf-court bookings — all covered. The user does almost nothing except the un-automatable parts like logging in.
Source: QbitAI
Industry
6. InfiniteCompute Closes a ¥2.3B Series C — Setting Dual Records for China’s AI-Infra Funding Race
InfiniteCompute (无问芯穹), an AI-native infrastructure provider, has closed its Series C at ¥2.3 billion (~$325M) in a single round, bringing cumulative fundraising to ¥4.3 billion — both single-round and cumulative records for China’s domestic AI-native infrastructure track. The investor roster spans three categories: state capital (Hangzhou High-Tech Jintou, the CICF Zhongguancun Fund, Zhuhai Sci-Tech, Xiamen Guoxing, Xuhui Capital, Futeng Capital, China Life Investment, China Development Bank Capital, etc.), strategic investors (Dabanc Holding, Lenovo, Jiukun, Tongge, Tencent, Baidu), and financial investors (Sequoia China, Hongtai, Huiyuan, Qiming, Dachen, Shunwei, Lishi). The company trains no models and ships no apps — it operates the layer between models and chips, optimizing efficiency and cost across many model algorithms and heterogeneous chip hardware to squeeze production-grade tokens out of every watt and every silicon cycle. Overseas, the same business has been repriced: Fireworks AI went from $4B to $17.5B in a year (+337%), Baseten +160%, Together AI +151% — all three backed by the same investor, Nvidia. SemiAnalysis founder Dylan Patel argues compute shortages will persist, and the agent boom is widening the gap by an order of magnitude: one agent task triggers dozens of chained inference calls versus a single chat turn.
Source: QbitAI
7. Huawei’s Hubble Invests in a ‘98-Generation’ PhD: Photonic-Crystal E-Skin for Better Robot Touch
Shenzhen-based Vista Robotics (视塔机器人) has closed an angel round with participation from Huawei’s Hubble Investment, Shenzhen High-Tech Investment’s funds, and Gaoxuan Investment, with Multidimensional Capital as sole financial advisor. Founder and CEO Zhang Wei, born in 1998, earned his PhD as a joint program between Southern University of Science and Technology (SUSTech) and City University of Hong Kong, spanning optical materials and robotics; the team averages about 23 years old, with core members from SUSTech’s Intelligent Liquid Crystal Optics Lab. The differentiation is materials-first: combining flexible photonic crystals with photosensitive elements to build e-skin under 1mm thick with sensing density above 10,000 points/cm² — attacking the thickness, flexibility, and large-area integration problems that hold back high-density tactile sensing. Robots can run, jump, and flip, but delicate manipulation — picking up an egg, tightening a nut — still lacks “feel”: vision tells a robot where an object is, but not how much force is applied after contact, whether it’s slipping, or how to adjust. Vista is building a full hardware-to-algorithm stack around tactile sensors, data-collection devices, and tactile-control models.
Source: QbitAI
8. Just 5 Years In, Chatbot Platform Character.AI Files for Chapter 11 Bankruptcy
Interactive chatbot platform Character.AI, which lets users create AI characters and converse with them, has filed for Chapter 11 bankruptcy following a slew of lawsuits and struggles to scale its popular platform. In court filings the company listed assets and liabilities between $50 million and $100 million. “Over the past year, Character.ai has taken meaningful steps to sustain and improve its business in the face of significant legal, regulatory, and operational headwinds. After a comprehensive review of the available alternatives, the board determined that this court-supervised sale process provides the Company with the best opportunity to maximize value for all stakeholders,” CEO Colin Adams said in a statement. The fall is a reminder that consumer-AI virality and litigation risk can move faster than revenue.
Source: TechCrunch
9. Amex Builds a Framework to Stop AI Agents From Making Unauthorized Purchases
Personal AI agents are increasingly making transactions on their users’ behalf — but LLMs are nondeterministic, so agents are prone to doing things they weren’t supposed to, raising the real possibility of an AI misfiring and paying for something it shouldn’t. American Express is introducing a framework designed to capture and authenticate a customer’s purchase intent before an AI agent completes the transaction. Rather than policing agents after the fact, the approach moves intent verification into the transaction flow itself — a template for how payment rails will need to evolve as agentic commerce scales.
Source: TechCrunch
Compiled from multiple media reports.









