Hugging Face releases tokenizers v1 with big speed gains

Hugging Face released a v1 release candidate of tokenizers focused on performance, saying it is often tens of times faster than v0.23. The team says tokenization can bottleneck large-scale training, high-concurrency serving and long inputs, leaving GPUs idle. IBM, NVIDIA and the ExecuTorch team contributed patches and cross-hardware testing.

• NVIDIA Launches DSX Ready for AI Factory Power, Cooling2026-09-21
• Nvidia Buys $1.5B More in SB Energy Ahead of IPO2026-09-21
• NVIDIA: Scaling physical AI demands safety at every layer2026-09-21
• Egypt's AI ecosystem reaches production scale2026-09-21
• NVIDIA: AI Security Is an Engineering Problem, Needs Full-Stack Controls2026-09-21