Overview

• Researchers in Beijing introduced SpikingBrain 1.0, a large language model from the Chinese Academy of Sciences’ Institute of Automation.
• The team says the model fires only necessary neurons and focuses on nearby context, cutting compute demands versus transformer-style global attention.
• According to the paper, SpikingBrain achieved comparable results using under 2% of the training data required by conventional models.
• Reported tests cite 25–100× speed improvements, with some ultra‑long sequences exceeding 100×, including a trial on a 4‑million‑token prompt.
• The system runs on China’s MetaX chip platform rather than Nvidia GPUs, with a preprint on arXiv and calls for independent benchmarking and peer review.