# TurboMamba speeds up training and inference, improves prediction

> The new TurboMamba architecture cuts prediction loss and runs up to 6× faster than comparable Mamba‑based models.

Oossa · 2026-10-16 · https://oossa.com/en/turbomamba-speeds-up-training-and-inference-improves-prediction

Researchers led by Hanxin Zhang introduced TurboMamba, a language‑model design that pairs a global Mamba‑3 block with a small Transformer for each token block. The combo adds three light tweaks: direct access to the previous block’s last token, a tiny score correction, and a learned mix of recent‑token frequencies. Tests on the public FineWeb‑Edu dataset show TurboMamba beats twelve baseline models, including standard Transformers and Mamba‑3 variants, in both speed and accuracy.

## The facts

- Mean test NLL is 9.4% lower than Mamba‑3 and training throughput is 2.60× higher.
- Batch latency speedup is 3.12× over Mamba‑3 for a 32‑sequence, 256‑prompt‑token workload.

## Why it matters

Researchers can train and run language models faster and cheaper, lowering the cost barrier for new AI projects.

## Sources & references

1. [TurboMamba: Faster Inference, Faster Training, and Better Prediction with Lexical Memory](https://arxiv.org/abs/2610.20682) – arXiv cs.LG, 2026-10-16

Last updated: 2026-10-16