v0.1.7

cheaper-inference/OmniRoutev0.1.7Mar 23, 2026by chonknick

Upgrading?

AI Summary

Replaced the HuggingFace tokenizers library with `tokie` to achieve massive performance improvements in reranking, reducing initialization time and dependencies while maintaining 100% accuracy.

Key Highlights

  • ~20-50x faster tokenizer initialization
  • ~260 fewer transitive dependencies
  • 100% token-accurate reranking results
  • No runtime dependency on HuggingFace Hub

New Features

  • Replaced `tokenizers` with `tokie`
  • Tokenizer served from cdn.mandex.dev/models/tokenizer.tkz

Full Release Notes

## What's Changed

- Replace HuggingFace `tokenizers` with `tokie` for reranking
  - ~20-50x faster tokenizer initialization (pre-built .tkz from CDN vs JSON parsing)
  - ~260 fewer transitive dependencies
  - 100% token-accurate, identical reranking results
  - No runtime dependency on HuggingFace Hub
- Tokenizer served from `cdn.mandex.dev/models/tokenizer.tkz` alongside reranker model

Related Repositories