cd /news/developer-tools/gigatoken-rb-a-ruby-port-of-gigatoke… · home topics developer-tools article
[ARTICLE · art-71351] src=github.com ↗ pub= topic=developer-tools verified=true sentiment=↑ positive

Gigatoken-rb – A Ruby port of gigatoken that's 1.6x faster

Gigatoken-rb, a Ruby port of the gigatoken BPE tokenizer, achieves 12 GB/s (2.78 billion tokens per second) on a Mac Studio M4 Max, running 1.6x faster than the upstream Python package on the same Rust engine. The gem is 340x faster than the fastest existing Ruby gem (tiktoken_ruby) and 1,050x faster than the tokenizers gem, according to benchmarks by developer Jetpks.

read4 min views1 publishedJul 24, 2026
Gigatoken-rb – A Ruby port of gigatoken that's 1.6x faster
Image: source

12 GB/s / 2.8 billion tokens per second in Ruby.

Ruby bindings for marcelroed/gigatoken, the fastest open-source BPE tokenizer around — running 1.6x faster than upstream's own Python package, on the same Rust engine.

Corpus MB/s (median) Gtok/s (median)
gigatoken-rb (this gem, Ruby)
11.9 GB 12,278
2.78
gigatoken (Python wheel, upstream) 11.9 GB 7,400 1.68
tiktoken (Python) 1.35 GB 69.7 0.0158
tiktoken_ruby 1.35 GB 30.7 0.0070
tokenizers gem (ankane) 1.35 GB 10.0 0.0023
tokenizers (Python, Hugging Face) 1.35 GB 5.6 0.0013

Mac Studio M4 Max, OpenWebText, GPT-2 tokenizer; every library produces the same tokenization. 340x faster than the fastest existing Ruby gem (tiktoken_ruby) and 1,050x faster than the tokenizers gem. Full methodology, exact counts, and the caveats that matter: docs/rb/benchmarks.md.

gem install gigatoken

Precompiled native gems ship for Apple Silicon macOS (arm64-darwin

) and x86_64/aarch64 Linux — on those platforms RubyGems grabs the binary automatically, no Rust toolchain, no compile wait. In a Bundler project it's one command:

bundle add gigatoken

(or drop gem "gigatoken"

into the Gemfile yourself).

Anywhere else (or with --platform ruby

to opt out of the binary), the extension builds from source. That needs a Rust toolchain: rust-toolchain.toml

pins the nightly, and rustup

fetches it automatically on first build.

require "gigatoken"

tok = Gigatoken::Tokenizer.load("openai-community/gpt2")

tok.encode("Hello, world!")             # => [15496, 11, 995, 0]
tok.decode([15496, 11, 995, 0])         # => "Hello, world!"
tok.encode_batch(["Hello!", "Another"]) # => [[15496, 0], [6610]]

tok.vocab_size                          # => 50257
tok.special_tokens                      # => {"<|endoftext|>" => 50256}

load

takes a tokenizer.json

path, a directory holding one, a HuggingFace Hub repo id, or a .tiktoken

mergeable-ranks file, and dispatches on shape. Hub downloads run over socketry's async-http

— no Python anywhere. Know what you have? Skip the dispatch:

Gigatoken::Tokenizer.from_file("tokenizer.json")
Gigatoken::Tokenizer.from_hub("openai-community/gpt2", revision: "main")
Gigatoken::Tokenizer.from_tiktoken("vocab.tiktoken")
Gigatoken::Tokenizer.from_json(File.binread("tokenizer.json"))

SentencePiece-BPE models (Llama, Gemma, Mistral — any tokenizer.json

with byte_fallback: true

) load through the same entry points and pick the right backend automatically. One difference: the SentencePiece core decodes text, so it validates input and raises Gigatoken::Error

on invalid UTF-8 instead of guessing.

encode_files

reads and tokenizes files entirely on the native side — document contents never materialize as Ruby objects. .gz

and .zst

decompress transparently.

tok.encode_files("owt_train.txt", separator: "<|endoftext|>")

jsonl   = Gigatoken::Native::JsonlFileSource.new(["docs.jsonl"], field: "text")
parquet = Gigatoken::Native::ParquetFileSource.new(["docs.parquet"], column: "text")
tok.encode_files(jsonl)

Pass packed: true

to encode_batch

or encode_files

and results land in a single IO::Buffer

of u32 token ids instead of a ragged Array of Arrays — no per-token Ruby allocation, the fastest way out of the engine:

packed = tok.encode_files("owt_train.txt", packed: true, separator: "<|endoftext|>")

packed.buffer       # => one IO::Buffer, every document's ids back to back
packed.lens         # => [12, 8, 41, ...] tokens per document
packed.token_count  # => total tokens
packed[3]           # => document 3's ids as an Array, on demand

encode_batch

and encode_files

release the GVL for the whole encode; the parallelism runs on the engine's rayon pool, not Ruby threads. Under Async

, give the fiber scheduler a worker pool (ASYNC_SCHEDULER_WORKER_POOL=true

) and the calling fiber yields to the reactor too. Design notes: docs/rb/async.md.

gigatoken bench openai-community/gpt2 owt_train.txt --doc-separator "<|endoftext|>"
gigatoken validate openai-community/gpt2 owt_train.txt --doc-separator "<|endoftext|>"

bench

reports MB/s and Mtok/s (--packed

for the fused packed path, --no-parallel

for the serial core). validate

confirms native split-and-encode agrees with a Ruby-side split through encode_batch

.

bundle install
bundle exec rake compile    # native extension (Rust nightly, via rust-toolchain.toml)
bundle exec rspec
bundle exec standardrb

The Ruby layer is fiber-first throughout — no Thread

, no Mutex

; all parallelism lives in the core's rayon pool. CI runs ubuntu + macos × Ruby 3.3/3.4/4.0, and release.yml

cross-builds the precompiled native gems (arm64-darwin, x86_64-linux, aarch64-linux).

This fork exists because I need fast tokenization in Ruby. The Rust core is changed as little as possible from upstream. Most of the python shell has been removed from this fork, but you can still find it upstream.

Not ported/no current plans:

  • the HF/tiktoken Python compat shims
  • padded-batch matrices
  • and BPE training

SentencePiece works but — matching upstream — is less optimized than the BPE path.

The engine is Marcel Rød's gigatoken. If it shows up in your research, cite that:

@software{roed2026gigatoken,
  author = {Marcel R{\o}d},
  title = {{G}igatoken: SIMD and Cache Hierarchies for 1000x Faster Byte-Pair Encoding Tokenization on Modern CPUs},
  url = {https://github.com/marcelroed/gigatoken},
  year = {2026},
}

AI Use Disclosure #

The Rust engine is upstream's — see upstream's AI-use disclosure for how that was built (majority hand-crafted, AI-assisted toward the end).

The Ruby port in this fork is 100% AI generated using Fable 5 and Sonnet 5 via space-architect over ~24 hours.

── more in #developer-tools 4 stories · sorted by recency
── more on @gigatoken-rb 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/gigatoken-rb-a-ruby-…] indexed:0 read:4min 2026-07-24 ·