# Exponentially Faster Language Modelling

arXiv: https://arxiv.org/abs/2311.10770 · HTML: https://mlsanity.com/paper/2311.10770/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Builds on 8 of its 11 references that are in the map.
- First posted on arXiv 15 Nov 2023.

## Builds on (most cited first, 8 of 8)

- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding](https://mlsanity.com/paper/1804.07461/) (2018)
- [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://mlsanity.com/paper/1910.01108/) (2019)
- [PaLM 2 Technical Report](https://mlsanity.com/paper/2305.10403/) (2023)
- [Patient Knowledge Distillation for BERT Model Compression](https://mlsanity.com/paper/1908.09355/) (2019)
- [Cramming: Training a Language Model on a Single GPU in One Day](https://mlsanity.com/paper/2212.14034/) (2022)
- [XtremeDistilTransformers: Task Transfer for Task-agnostic Distillation](https://mlsanity.com/paper/2106.04563/) (2021)
- [Fast Feedforward Networks](https://mlsanity.com/paper/2308.14711/) (2023)

## Built on by (most cited first, 4 of 4)

- [Deploying Foundation Model Powered Agent Services: A Survey](https://mlsanity.com/paper/2412.13437/) (2024)
- [Dynamic sparsity in tree-structured feed-forward layers at scale](https://mlsanity.com/paper/2604.08565/) (2026)
- [HASTE: A Framework for Training-Free, Dynamic, and Steerable Compression of Pre-Trained Convolutional Neural Networks](https://mlsanity.com/paper/2606.30516/) (2026)
- [The SpiNNaker2 chip: a many-core platform for flexible and scalable brain-inspired computing](https://mlsanity.com/paper/2607.24396/) (2026)

## Abstract (arXiv)

Language models only really need to use an exponential fraction of their neurons for individual inferences. As proof, we present UltraFastBERT, a BERT variant that uses 0.3% of its neurons during inference while performing on par with similar BERT models. UltraFastBERT selectively engages just 12 out of 4095 neurons for each layer inference. This is achieved by replacing feedforward networks with fast feedforward networks (FFFs). While no truly efficient implementation currently exists to unlock the full acceleration potential of conditional neural execution, we provide high-level CPU code achieving 78x speedup over the optimized baseline feedforward implementation, and a PyTorch implementation delivering 40x speedup over the equivalent batched feedforward inference. We publish our training code, benchmarking setup, and model weights.