SpikingBrain-1.0 Achieves 100× Faster Inference with Brain‑Inspired Spiking Architecture

Transformer‑based large models dominate AI research, but their training cost grows quadratically with sequence length and inference memory grows linearly, limiting the handling of ultra‑long sequences.

In contrast, the human brain processes perception, memory, language and reasoning with only ~20 W power, prompting researchers to explore brain‑inspired alternatives that could break Transformer bottlenecks.

Building on this motivation, the Institute of Automation, Chinese Academy of Sciences and the National Key Lab of Brain‑Inspired Intelligence proposed an “endogenous‑complexity” architecture. The design formally connects spiking‑neuron intrinsic dynamics with linear‑attention mechanisms, showing that conventional linear attention is a special case of dendritic computation. The team released the open‑source SpikingBrain‑1.0 model, a spiking‑neuron‑based foundation model with linear and mixed‑linear complexity, together with a Triton operator library, model‑parallel strategies and cluster‑communication primitives for domestic GPU clusters.

Experimental validation shows four breakthroughs: (1) efficient training with very few data, (2) inference speedup by orders of magnitude, (3) a domestically controllable brain‑inspired model ecosystem, and (4) a dynamic‑threshold multi‑scale sparse mechanism achieving 69.15% sparsity and low‑power operation. Notably, SpikingBrain‑7B reduces the Time‑to‑First‑Token on a 4 M‑token sequence by more than 100×, runs stably for weeks on hundreds of MetaX C550 GPUs, and reaches 23.4% FLOP utilization.

This is the first large‑scale brain‑inspired linear foundation model announced in China and the first to run training and inference on a domestic GPU‑compute cluster. Its ability to process ultra‑long sequences makes it promising for legal and medical document analysis, complex multi‑agent simulation, high‑energy‑physics experiments, DNA‑sequence analysis, and molecular‑dynamics trajectory modeling.

The model and its tutorial are hosted on the HyperAI website. Users can launch a one‑click deployment by:

Opening the HyperAI homepage, selecting the “Tutorial” page, and choosing “SpikingBrain‑1.0 Based on Endogenous Complexity”.

Clicking “Clone” to copy the tutorial into their own container.

Selecting an NVIDIA RTX A6000 48GB GPU with a PyTorch image and choosing a billing plan (pay‑as‑you‑go or subscription). New users can claim free RTX 4090 and CPU time via the invitation link.

Waiting ~3 minutes for resource allocation, then opening the demo via the provided API address after identity verification.

Entering queries in the chat window to interact with the model.

As a demonstration, the author asked the model to generate a CSS/JavaScript snippet for a sticky header; the returned code is shown in the accompanying screenshot.

Overall, SpikingBrain‑1.0 showcases a viable path toward energy‑efficient, ultra‑long‑sequence AI by marrying spiking neuroscience with modern linear‑attention architectures.