AI Daily — 2026-10-08

English 中文

Cua launched open-source computer-use agent tools, Swift LLMs hit 2.2M downloads with free compute, and one person AI-hacked South Korean banks.

Covering 22 AI news items

🔥 Top Stories

1. Cua Opens Computer-Use Agents with Open-Source Drivers and Benchmarks

Cua has released an open-source stack that gives AI agents full desktops on Macs and other owned machines, including Cua Driver for desktop automation, Lume for local VMs, CUA-S1 decision models, and Cua Bench. By packaging drivers, VMs, models, and evaluation in one platform, it lowers the barrier to building and benchmarking computer-use agents across operating systems. Source-github

2. Swift Open-Source Reasoning LLMs Surpass 2.2M Downloads

UkisAI’s Swift models have crossed 2.2 million downloads by reducing token usage 58.3% and improving speed 1.95x without accuracy loss, largely by penalizing pathological overthinking. The team also announced early access to new models and free compute for researchers, strengthening the case for efficient reasoning models in local and production deployments. Source-reddit

3. Single Attacker Breaches South Korean Banks with Open-Source AI Stack

South Korea’s largest banks were reportedly hit by a cyberattack carried out by one individual using the open-source AI penetration tool ARTEX alongside DeepSeek v4.1-Flash, GLM-5.3, Grok 4.6, and Claude Code, according to CrowdStrike. The incident highlights how accessible AI tooling can lower the expertise and resource barriers for high-impact cyberattacks, raising pressure on defenders to adopt AI-assisted security workflows. Source-reddit

📰 Featured

Efficient Inference & Hardware

STEPQuant: Error-Aware Quantization for Recurrent States in Linear Attention — STEPQuant differentiates quantization error by temporal and spatial dimensions in recurrent states, enabling lower-precision linear-attention serving without severe accuracy degradation. Source-huggingface
$2,800 8x Radeon Pro V620 Rig Runs LLM via Custom vLLM Fork — A 256 GB VRAM build using eight Radeon Pro V620 cards reached 60–100 tokens/s decode and 3000+ tokens/s prefill after a custom vLLM fork addressed RDNA2 kernel limitations. Source-reddit

Multimodal & Generative Research

GRACE: Generation-aware Latent Compression for Efficient Video Generation — GRACE reduces tokens processed by the video diffusion transformer through generation-aware latent compression, targeting reconstruction quality, channel growth, convergence slowdown, and pretrained-latent mismatch. Source-huggingface
UltraText Bench: Bilingual Benchmark for Visual Text Rendering in Image Generation — UltraText Bench offers 432 human-reviewed prompts across 24 real-world scene categories and three difficulty levels, split evenly between English and Chinese. Source-huggingface
VepAgent Uses Tool-Augmented RL for Video Event Prediction — VepAgent combines causal-transition reasoning with tool-augmented reinforcement learning to help multimodal LLMs predict unobserved causal transitions in video. Source-huggingface
UniWAM Integrates Physical Reasoner, World Generator, and Action Predictor — UniWAM unifies semantic VLA reasoning with spatiotemporal world-model priors through a physical reasoner, world generator, and action predictor. Source-huggingface

AI Security & Agent Tooling

Cloudflare Releases Security Audit Skill for Coding Agents — Cloudflare’s security-audit-skill runs structured six-phase audits with isolated agents, validation, and independently verified findings for fleet-wide vulnerability discovery. Source-github

⚡ Quick Bites

Six AI Decision Models Play Pac-Man, Open-Source Leaderboard Released — Six AI decision models compete in Pac-Man with an open-source leaderboard for reproducible agent evaluation. Source-reddit
RTX 4090 Benchmarks Four Open Decision Models — A local RTX 4090 test benchmarks four open decision models for performance and decision quality. Source-reddit
audio.cpp Update: Higgs TTS 48% Less VRAM, Models Faster — audio.cpp updates reduce Higgs TTS VRAM usage by 48% while speeding up multiple supported models. Source-reddit
JetBrains Releases Mellum2.1 Small MoE Thinking GGUF Model — JetBrains adds a small MoE thinking GGUF model to its Mellum2.1 collection for local reasoning workloads. Source-reddit
Saluki 27B Claims 96% of Qwen 3.8 Performance at 1/7 Size — Saluki 27B claims roughly 96% of Qwen 3.8’s performance at one-seventh the model size. Source-reddit
LittleBit Enables Ultra Low-Bit Quantization via Latent Factorization — LittleBit introduces latent-factorization quantization for ultra-low-bit LLM deployment. Source-reddit
Turkish TTS Trained from Scratch with Drifting Method on RTX 5090 — A Turkish TTS model was trained from scratch using the drifting method on a single RTX 5090. Source-reddit
Strata rewrote GitHub history to erase Claude co-author evidence — Strata reportedly rewrote its GitHub history to remove evidence that Claude co-authored changes. Source-reddit
Nvidia CMP 170HX Cards Reportedly Unlock 40GB to 48GB VRAM — Owners report unlocking Nvidia CMP 170HX cards from 40 GB to 48 GB VRAM for AI tasks. Source-reddit
Niche Hugging Face Finetune VeriLoop-E2 Shows Success in Codebases — VeriLoop-E2, a niche Hugging Face finetune, is reportedly showing high success in codebase tasks. Source-reddit
Waterblock Released for AMD R9700, Enabling Quiet 4x/6x AI Builds — A new waterblock for the AMD R9700 enables quieter 4x/6x GPU AI builds. Source-reddit
llama.cpp Creator ggerganov Presents on Stage at Event — llama.cpp creator ggerganov presented on stage, underscoring continued momentum for local inference. Source-reddit

Generated by AI News Agent | 2026-10-08