Tencent's chief AI scientist Yao Shunyu warns that AI will soon be capable of developing bioweapons, and argues that neither fully open nor fully closed models can address the risk alone. He advocates a hybrid open-closed approach so all parties can use the technology for defense, urging the industry to proceed carefully. Tencent also previewed an open-weight model, Hy4, with a safety report claiming its biological capabilities fall below the high-risk threshold.
Open-weight
开放权重
A release approach that publishes model weights for downstream deployment and fine-tuning.
Curated Intelligence
Thinking Machines, the startup founded by former OpenAI CTO Mira Murati, has released Inkling 2 Small, a 594-billion-parameter mixture-of-experts model that activates only 17 billion parameters per inference. It scores 31 on the Artificial Analysis Intelligence Index, an 18-point gain over its predecessor and roughly on par with Gemini 3.8 Flash, with notable improvements on agentic tasks, though cost per task is six times higher. The model debuts as a proprietary offering on the Tinker platform, with Apache-licensed weights to follow.
Ricursive Intelligence, founded by former Google researchers, released Ripple-65B Preview, a dense open-weight model pretrained on 12T tokens with 1M context that it claims outperforms the much larger Meta Muse Spark Open 816B and NVIDIA Nemotron 3 Ultra 550B on TerminalBench 2.1. It is a general-purpose variant of the firm's chip-design model, with self-reported training cost of about $32 million.
Verne Robotics released V0.5, an open-weight world action model built on the Wan2.2 video generator and trained on over 1.7M demonstrations, predicting the next second of arm, gripper and mobile-base motion. It learns new industrial tasks from under two hours of demos and degrades far less than Physical Intelligence's π0.5 under clutter and lighting shifts.
Ricursive Intelligence releases Ripple-65B Preview, a general-purpose open-weight model pretrained on 12 trillion tokens. Achieving 80.52 on TerminalBench 2.1, it surpasses larger models like Reflection Beam to set a new US open-weight state-of-the-art in agentic coding.
Moonshot AI has released Kimi K3.1 Pro, an open-weight MoE model with 2.8T total and 104B active parameters. It features a 1M-token context window and multimodal capabilities, demonstrating significant improvements in long-horizon coding, agentic tasks, and financial modeling with a 65% reduction in hallucinations. The model is now available via OpenRouter.
Moonshot AI has released Kimi K3.1 Pro, its most capable open-weight model to date, featuring a 1-million-token context window and substantial improvements in coding, 3D creation, and financial research. A Highspeed variant is available now via API, with full model weights scheduled for release on November 4.
Meta has released Muse Spark Open, an 816-billion-parameter multimodal model with a 1M-token context window, claiming top performance among US open-weight models. Designed for coding and agentic tasks, it natively supports Apple Silicon but requires at least 1TB of RAM for local execution.
Meta has released Muse Spark Open, an 816-billion-parameter open-weight multimodal model capable of processing text, images, video, and audio with a 1-million-token context window. It performs on par with leading Chinese open-source models like Qwen and GLM in benchmarks and is specifically optimized for local execution on Apple's M5 Ultra chip.
OpenAI has released Privacy Filter, an open-weight small model designed for high-throughput detection and redaction of Personally Identifiable Information (PII). Capable of running locally to ensure data privacy, it performs context-aware masking in a single pass. The release is currently in preview, aiming to gather feedback from the research community to refine performance.