NVIDIA Expands Nemotron Model Suite with New MoE Release
NVIDIA has announced the release of its Nemotron Bolt 3.5 74B A9B model, effective October 15. According to reports circulating on Reddit, the new model utilizes a Mixture-of-Experts (MoE) architecture, featuring 74 billion total parameters with 9 billion active parameters.
The model is designed to fit within a 24GB memory footprint through the use of 2-bit quantization. Developers indicate that advanced quantization-aware training techniques were employed to maintain model quality despite the compression. The technical objective of the Bolt 3.5 74B A9B is to provide a balance of speed and intelligence, aiming to outperform popular dense models in velocity while exceeding the intelligence of similar MoE models within the same memory constraints.
NVIDIA is also scheduled to release the Nemotron 3.5 Bolt E4B 4GB model on October 23.
Robotis and NVIDIA Partner for Humanoid Mass Production
At the Humanoid Summit Seoul 2026, held at COEX, Robotis announced a partnership with NVIDIA to initiate mass production of its “AI Sapiens” humanoid robots. The company has established a production target of 10,000 units annually.
Kim Byung-soo, founder and CEO of Robotis, stated that the company aims to establish humanoids as a versatile platform capable of performing multiple functions, similar to the evolution of the smartphone. To support this, Robotis plans to release a lineup of approximately 70 actuator models starting early next year.
During the summit, the AI Sapiens platform demonstrated stability in boxing and squatting motions. Kim noted that the robot, which stands approximately 136 cm tall and weighs a similar amount to Unitree’s G1, offers more than double the battery operating time due to higher actuator efficiency.
The company’s strategy involves a “data pyramid” approach, utilizing NVIDIA’s Cosmos Transfer to resolve synthetic data issues, such as shadow misidentification. Robotis intends to combine real-world data collection with training across thousands of virtual environments to improve learning efficiency. NVIDIA also presented its GR00T 1.7 general-purpose robot intelligence model and a roadmap for the upcoming GR00T 2.
Global AI Infrastructure and Competition
As the AI sector expands, competition in hardware and specialized chips continues to evolve. Industry discussions on Reddit suggest that while Microsoft is looking to increase its reliance on homegrown AI chips to reduce dependence on NVIDIA, the transition to AI-specialized ASICs may take several years as the industry awaits the stabilization of the “model wars.”
In China, significant investments are being directed toward space-based AI computing networks and domestic chip development. Recent reports indicate that China has launched eight new satellites to advance its space-based AI infrastructure. Additionally, companies such as Baidu have reported the successful operation of large-scale clusters, including a cluster of 30,000 self-developed P800 Kunlun chips.
Other regional developments include Alibaba’s launch of the open-source Qwen3 model and Kuaishou Technology’s development of video generation tools intended to compete with OpenAI’s Sora.