Cua just open-sourced CUA-S1-FORMS, the first in its family of small, specialized models built for computer use. Instead of clicking through fields one step at a time, it processes entire forms in a single pass using four core actions and field values pulled from a document. The release is MIT-licensed with the model, dataset, and code all public.
Alexandr Wang just launched Muse for Mac, an AI agent capable of handling tasks across your files, messages, calendar, and notes directly on your computer. You keep full control over data access, and it requires explicit confirmation before executing any sensitive actions.
prism-ml just launched Bonsai-2, a collection of radically compressed models that restrict weights to -1, 0, and +1. The approach shrinks a 27B-parameter base model to under 6GB—slashing its footprint by over 9x while preserving baseline performance on consumer-grade GPUs and Apple silicon.
Cactus just rolled out Needle 3, a tiny foundation model packed into a single 8–29 MB binary for smartwatches and microcontrollers. It trades general chat capacity to excel at offline tool calls and data extraction, scaling from 2 to 20 layers based on hardware limits. Why it matters: developers can now build instant, fully offline actions directly on tiny edge devices without relying on the cloud.
Perplexity just launched Portable Computer on Windows, bringing its multi-step agent to NVIDIA GeForce RTX PCs and RTX PRO workstations with 24GB+ VRAM. It runs models like Qwen 3.8 27B entirely on-device, letting you analyze files and connect to tools like Slack and GitHub without leaking data or burning cloud credits.
Apple just rolled out Siri AI alongside the next generation of Apple Intelligence, rebuilding its assistant to tap into personal context across messages, mail, and on-screen content. It can now take multi-step actions across third-party apps and sync chats across devices in a dedicated Siri app. The beta is live today in English.
Zhongguancun researchers just released ZGCM-1, an open 7.39B model specialized in mathematical reasoning and agentic web search with 256K context support. It hits 97.13% on MATH-500 and comes with fully disclosed data and training pipelines ready for consumer GPUs. Why it matters: Small, highly specialized open-source models are becoming genuinely capable at complex reasoning tasks without enterprise hardware.
Edge0 just open-sourced a streaming MoE inference framework that treats storage like memory, running 35B models in roughly 2.5 GB of active RAM on iPhones and Macs. By predicting routing one step ahead to hide SSD read latency, it squeezes large models onto low-memory consumer devices without crushing performance.
Apple just confirmed that its delayed Siri AI assistant will launch in beta on September 14 alongside iOS 27, complete with daily server usage caps and plans for future paid tiers. The initial release only supports English, blocks users under 13, and is unavailable in mainland China. Why it matters: Apple's personal assistant overhaul is finally arriving, but cloud compute constraints mean free, unlimited access is officially off the table.
Apple just unveiled the iPhone Duo, its first foldable device featuring a 7.6-inch inner display and the A20 Pro chip. The dual-screen flagship runs iOS 27 and introduces an overhauled Siri AI capable of onscreen awareness and real-time AI camera tracking. It marks Apple's biggest hardware leap yet to anchor personal AI across flexible displays.
Apple just introduced the iPhone 18 Pro lineup, packed with a 2nm A20 Pro chip that doubles on-device AI power via a 32-core Neural Engine. The flagship adds a hardware variable aperture camera and deeply integrates an all-new Siri AI. Pre-orders kick off September 12.
Tencent just open-sourced AuK, a 1.5B foundation model that lets you generate and edit speech using simple text prompts. Beyond zero-shot voice cloning, it can seamlessly insert or delete words from existing audio, tweak emotional tone, and remove background noise. It brings text-editor simplicity to voice workflows and runs on consumer GPUs.
OpenBMB just open-sourced MiniCPM5-2B, a compact 2-billion-parameter dense model that punches above 4B-class models like Qwen3.5-4B across coding and math benchmarks. Weighing only around 5GB, it easily runs locally on consumer edge hardware, and the team also released the full high-quality dataset behind it.
Rivian just rolled out its 2026.31 OTA update, bringing RivianOS 2 to all R1 models in the vehicle's biggest software overhaul to date. The update unifies R1 and R2 onto a single AI-defined platform with an overhaul of the UI, Unreal Engine real-time cabin visuals, and new granular privacy toggles.
Nvidia just released Personal AI Router (PAIR), a free open-source tool that links idle computers across your home network into a shared local AI compute cluster. It works with RTX 20-series GPUs and newer, plus Apple M4 chips, automatically stepping aside when you start gaming. A beta is out now for Windows, Linux, and macOS.
H3 World just turned the open-source Miniax H3 video generator into an interactive game engine. Instead of building a complex control rig, it simply translates keystrokes into short text prompts that guide the video's motion in real time. The project code, local setup instructions, and training scripts are now fully open-sourced.
Perplexity just launched PII-TRACE and PII-Tracer, a benchmark and 0.6B local model built on Qwen3 to screen sensitive data before routing tasks to cloud models. It achieved the highest character F1 among 12 detectors and stayed accurate across multi-turn chats. Why it matters: hybrid AI only works if on-device gates know exactly what to hide.
Alibaba just released Qwen-Drive-1.0, a vision-language foundation model tailored for autonomous driving. Built on Qwen3.5-4B without changing the underlying architecture, it plugs in a BEV perception head and a diffusion planning expert to handle 3D sensing and trajectory generation. It shows general VLMs can take the wheel without rebuilding them from scratch.
Google Research just open-sourced TimesFM 3.0, its third-generation time-series foundation model that now natively handles multivariate forecasting. It evaluates multiple correlated trends and planned future events with zero-shot accuracy, runs on just 1GB of VRAM, and completes full-horizon forecasts in a single forward pass under a non-commercial license.
NVIDIA just rolled out DLSS 4.5 Ray Reconstruction, powered by a second-generation transformer model to dramatically sharpen ray-traced and path-traced graphics. The neural rendering update is already live across 30 games via the NVIDIA app, cutting down ghosting artifacts. NVIDIA also previewed RTX Mega Geometry, bringing full-detail Nanite ray tracing to upcoming titles.
