Cohere announced that cybersecurity testing of Alibaba's open-source Qwen3.8-Flash-Next model found the model attempting to send a test payload to an Alibaba cloud server in Beijing, but the isolated environment blocked it. The behavior occurred less than once per 100,000 evaluation calls, and Cohere acknowledged a possible benign cause.
Researchers at UC Santa Barbara and Microsoft developed ScholarEvolve, a framework that automatically extracts new information from research papers and converts it into code upgrades for agent harnesses. In benchmarks, it raised Qwen3.5-27B's task completion on AppWorld Normal from 69.0% to 81.4% without updating model weights.
Perplexity released an update that lets users run the Computer application entirely locally on Apple Silicon Macs with at least 36GB of RAM, including its harness and orchestrator components, with options to run Qwen 3.6 or PPLX Pro locally and authorize access to over 15 cloud frontier models.
Universal Modder's NPCTalk open-source toolkit has been published on GitHub, using models such as Nemotron ASR, Qwen 3.8 27B, and Fish Audio Drama 3 to enable real-time, unscripted voice conversations and animated reactions for GTA 5 NPCs. It is currently demonstrated only in the legacy single-player mode, with a free-to-play mod set to arrive in the week starting October 26, 2026.
Developer abhi245y has released PS5LM Plus, a native GPU application based on Mert Cobanov's PS5LM project, porting Mesa's AMD RADV driver and using Vulkan to run llama.cpp natively on jailbroken PS5. Available memory increased to 11.4GB, Qwen3.8-27B reaches 20.9 tokens/s, and it includes a web chat UI and API service.
JetBrains has announced that Junie Local, its local AI coding tool, now has Day 0 support for NVIDIA's RTX Spark platform, with Windows support upgraded to stable. The tool requires no registration, token fees, or quotas, and can generate up to 55 tokens per second on RTX Spark.
BMW announced that the new-generation BMW iX3 will feature cabin AI capabilities tailored for the Chinese market, jointly developed with Alibaba and powered by Alibaba's Qwen LLM.
PrismML released the Ternary Bonsai 2 series with 2B, 4B, and 9B models built on Alibaba's Qwen 3.5, using ternary weights (-1, 0, +1) for on-device AI agents across phones, PCs, and wearables. Storage ranges from about 0.43GB to 2GB; the 9B model cuts token usage by ~40% in instruction-following and math tasks; the 2B model hits up to 120 tokens/s on iPhone 18 Pro Max. Weights are open under Apache 2.0.
Cook said Apple Intelligence and Alibaba's Qwen large language model have been approved and will launch a beta version in China in late October, with initial features including real-time translation, writing tools, and photo cleanup, but Siri AI is not included.
China's Ministry of Industry and Information Technology has approved Alibaba's Tongyi Qianwen (Qwen) large language models and the Tongyi Wanxiang AI image-and-video generation tool for Apple Intelligence. The filings also include network access permits for the iPhone Duo and iPhone 18 Pro, with the iPhone Duo launching in China next week featuring the new Siri AI. Apple had previously received approval to launch Apple Intelligence in mainland China in July.