GoKawiil
Tech News
Meta and Apple have struck a partnership allowing developers to run Meta's open-weight AI models locally on Mac hardware. The collaboration gives developers an alternative to relying on cloud-based computing power for running advanced AI systems.
Nvidia has launched Nemotron 3.5 Super VL, a 120-billion-parameter open-weight model aimed at agentic enterprise tasks. The company says benchmark testing by Artificial Analysis shows it is five times faster than the next-best open model in its class, outperforming InclusionAI's Ling 3.0 Flash and DeepSeek's V4.1 Flash. Nvidia also cited a 44% accuracy gain in a medical insurance claims task after post-training, surpassing Claude Opus 5 on that specific benchmark.
Bloomberg reports that open-weight AI models, which can be run locally and stripped of safety guardrails, are increasingly being used to produce child sexual abuse material. According to the report, these models can be trained on existing CSAM and generate new content offline, making detection by law enforcement difficult, and have spawned large dark web communities around this practice. Child safety expert Rebecca Portnoff told Bloomberg that over 500 open source models are now implicated, up from fewer than a dozen two years ago, while NCMEC-reported incidents of AI-generated CSAM jumped from 67,000 to more than 1.5 million in a year.
Mozilla's latest State of Open Source AI report, using data through September 1, finds top Chinese open-weight models are closing in on closed US frontier systems, lagging by only a few percentage points on key benchmarks while costing far less to run. The report estimates the capability gap at roughly 4.4 months based on METR task-horizon data, close to Epoch AI's independent four-month estimate, and notes one Chinese model scored within a point of a leading Claude model on Terminal-Bench at under a fifth of the price.
Tom's Hardware benchmarked Alibaba's newly released Qwen3.8 27B open-weight AI model across discrete GPUs like the RTX 5090, RTX 4090, RTX 3090, and AMD/Intel cards, as well as unified-memory systems including the DGX Spark, Mac Studio, and Ryzen AI Halo. Despite the four-bit quantized model needing only about 17GB of VRAM, the tests found that software stacks and inference engines often bottleneck real-world performance, affecting time-to-first-token and throughput once context windows fill up.
AI language models fall into three broad categories based on how much of their inner workings are shared publicly: proprietary systems that keep code and weights closed, open-weight models that release trained parameters without full training data or code, and open source models that make the entire stack available for inspection and modification. The distinctions affect how much control, transparency, and customization developers and businesses get when building on top of these systems.