Q Labs reports 4000x speedup in open-source NOiSE training project
Time to hit the target loss fell from 1.76 hours to 1.6 seconds, and from October 22 the evaluation expands to 30 million tokens, according to aitimes.com.
Q Labs announced on October 22, 2026, that its open-source NOiSE project has recorded a roughly 4000x speedup, cutting the time needed to reach the target loss from 1.76 hours to 1.6 seconds, according to aitimes.com. The company also said it has ended the 10-million-token evaluation phase and increased training data to 30 million tokens.
The experiment trained a small GPT model with 37.7 million parameters on up to 10 million FineWeb tokens in a single pass, with the goal of lowering validation loss to 4.2. Within ten days the best loss improved from 4.167 to 3.676, below the 3.769 baseline produced by backpropagation using the Adam optimizer.
Q Labs developed the backpropagation-free training algorithm Dust and opened a public experiment based on it. In the new phase, evaluation runs on a single node with 8 H100 GPUs against a target loss of 3.5, and a track has been added that allows free changes to model architecture.
Backpropagation still ahead in like-for-like tests
According to aitimes.com, comparisons that applied the same optimization algorithm to both sides showed that backpropagation remained faster. In its original research, Dust describes changing activation values instead of weights and estimating weight update directions from those results.
Quick answers
What speedup did Q Labs report for NOiSE?
Q Labs reported a roughly 4000x speedup, with the time to reach the target loss falling from 1.76 hours to 1.6 seconds within ten days.
What are the settings for the expanded NOiSE evaluation from October 22, 2026?
It uses 30 million training tokens on a single node with 8 H100 GPUs and a target loss of 3.5, and adds a track allowing free changes to model architecture.
Is Q Labs' backpropagation-free approach faster than backpropagation?
In comparisons applying the same optimization algorithm to both sides, backpropagation remained faster, according to aitimes.com.