Hugging Face ships tokenizers v1, often tens of times faster than v0.23

ariG23498 · x · 2026-09-21

Hugging Face released a release candidate of tokenizers v1 focused on performance. As models get faster and workloads scale — massive training sets, high-concurrency serving, repeated long-input processing — CPU tokenization can starve the model and leave GPUs idle. The v1 refactor delivers benchmarks often tens of times faster than v0.23, measured across single-threaded, multi-threaded and thread-scaling modes. NVIDIA, IBM and the ExecuTorch team contributed patches and hardware testing; the team credits prior open-source tokenizer work (tiktoken, kitoken, etc.) for many of the ideas.

Related event: Hugging Face Ships tokenizers v1, Up to 30x Faster(8 posts)→

Original post →

More from Infra

Infra channel →