As of September 2025
Most large language models burn energy across millions of parameters that have nothing to do with the question in front of them. A spiking neural network does the opposite: neurons fire only when something relevant happens, and the rest stay silent.
The SpikingBrain technical report, out of a Chinese group in September 2025, puts hard numbers on what that buys. They are real figures, as long as you keep the conditions attached to each one.
Long-context speed
>100x
time-to-first-token speedup at 4M-token context versus a Transformer baseline (SpikingBrain report, Sep 2025).
MAC energy
97.7%
energy reduction in multiply-accumulate operations versus FP16, from sparse spiking activation.
Training data
<2%
of from-scratch training data - about 150 billion tokens via conversion-based continual pre-training.
What is verified
Two models, SpikingBrain-7B and a 76B mixture-of-experts, were trained and run on MetaX hardware with no NVIDIA chips in the loop. The 76B reaches performance comparable to Llama2-70B and Mixtral-8x7B, and the report measures about 69 percent activation sparsity, so most of the compute never fires.
What the headlines hide
The eye-catching figures carry conditions. “Over 100x faster” is time-to-first-token at a 4-million-token context, not general throughput. “97.7 percent less energy” is per multiply-accumulate operation against FP16, not wall-clock power draw. The efficiency is real, but it lives in long-context inference and sparse arithmetic, not everywhere at once.
The brain solved this millions of years ago - spend energy only where the work is. The interesting part here is doing it on hardware everyone assumed was inadequate.
