OlexwebThe facts, in its own words
- Highlights LLM tokens as you type using only OpenType font features“A monospace font that highlights LLM tokens as you type. No JavaScript, just OpenType substitution rules and a color table.”, george.mand.is
- Includes 28 styles, each mapped to a specific tokenizer model“It's one big typographic family—"Tokenizer Mono" based on IBM Plex — with a style per tokenizer model. Font Book on macOS should register them as one family with twenty-eight styles.”, george.mand.is
- Covers tokenizers for OpenAI, Meta, Google, Mistral, Alibaba, DeepSeek, and more“Vendor Face Used by OpenAI cl100k OpenAI GPT-4, GPT-3.5, text-embedding-3; also Phi-4, OLMo 2, IBM Granite 4 o200k OpenAI GPT-4o, o1, o3, GPT-4.1, gpt-oss p50k OpenAI Codex, text-davinci-002/003 r50k OpenAI GPT-2, GPT-3 clip OpenAI CLIP, Stable Diffusion prompts (lowercases; no whitespace tokens) Meta llama4 Meta Llama 4 llama3 Meta Llama 3 / 3.1 / 3.2 / 3.3; also SmolLM 3 llama2 Meta Llama 2 (SentencePiece) Google gemma3 Google Gemma 3 gemma Google Gemma 1 / 2 t5 Google T5 (SentencePiece Unigram, not BPE) bert Google BERT (WordPiece; ## continuations approximated) Mistral mistral Mistral 7B v0.3 (SentencePiece) tekken Mistral Nemo, Mistral Small 3 (Tekken); also NVIDIA Nemotron-H, Swiss AI Apertus Alibaba qwen Alibaba Qwen 2 / 2.5 / 3 DeepSeek deepseek DeepSeek V3 / R1 Zhipu glm Zhipu GLM-4.5 Tencent hunyuan Tencent Hunyuan Baidu ernie Baidu ERNIE 4.5 (SentencePiece) ByteDance seed ByteDance Seed-OSS 01.AI yi 01.AI Yi (SentencePiece) IBM granite IBM Granite 3 and Granite Code (StarCoder vocabulary); also StarCoder TII falcon3 TII Falcon 3 falcon TII Falcon EleutherAI neox EleutherAI GPT-NeoX, Pythia BigScience bloom BigScience BLOOM (multilingual) UW superbpe UW SuperBPE for OLMo 2 (tokens cross word boundaries) Moonshot kimi Moonshot Kimi K2 (tiktoken-format vocabulary)”, george.mand.is
- Source code is publicly available on GitHub“The code is at: github.com/georgemandis/tokenizer-mono .”, george.mand.is
- Works without JavaScript — highlighting is drawn by the font itself“You can play with it below. Every highlight is drawn by the font. No scripting! The only JavaScript here is swapping which face is loaded.”, george.mand.is
- Does not include tokenizers for Anthropic Claude or Google Gemini“Anthropic's tokenizer isn't public, as far as I can tell, so there's no Claude face. Neither is Gemini's, though I wouldn't be shocked if it's some evolution of Gemma?”, george.mand.is
- Based on IBM Plex monospace font family“It's one big typographic family—"Tokenizer Mono" based on IBM Plex — with a style per tokenizer model.”, george.mand.is
Free. Free to use and open source “The code is at: github.com/georgemandis/tokenizer-mono . You can play with it below.”
- AI developers comparing tokenizer behavior across models
- Researchers studying tokenization differences visually
- Students learning how LLMs split text into tokens
- Prompt engineers optimizing token usage for cost and context
Up and running today
- 01Download the font from github.com/georgemandis/tokenizer-mono
- 02Install the font family on your system (28 styles install as one family)
- 03Open any editor with OpenType feature support (VS Code, Sublime Text, Figma, etc.)
- 04Select 'Tokenizer Mono' and choose the style matching your target model (e.g., cl100k for GPT-4)
- 05Type text and watch token boundaries highlight automatically
Tokenizer Mono is a clever, zero-dependency way to see tokenization live in your editor — great for quick comparisons and building intuition. It's not a replacement for precise token counting (the font approximates some behaviors), and it lacks Claude/Gemini tokenizers, but for OpenAI, Llama, Gemma, Mistral, and others it's an instant, install-once tool that works everywhere fonts work.
- Instant visual tokenization in any OpenType-aware editor
- Side-by-side comparison of how different models tokenize the same text
- Learning tokenization concepts without leaving your code editor
- Offline use — no internet or API calls needed
- Approximates some tokenizer behaviors (e.g., BERT WordPiece continuations)
- No styles for Anthropic Claude or Google Gemini tokenizers
- Requires an editor with full OpenType GSUB support for highlighting to appear
- Newer models like GPT-6 (Astra) not yet covered
Is Tokenizer Mono Font free?
Yes. The font is free and open source — the code is on GitHub and you can try it directly on the project page.
How do I use Tokenizer Mono Font?
Install the font from GitHub, then select 'Tokenizer Mono' in any OpenType-aware editor and choose the style matching your target model (e.g., cl100k for GPT-4). Token highlights appear automatically as you type.
Which tokenizers does Tokenizer Mono support?
It includes 28 styles covering OpenAI (cl100k, o200k, p50k, r50k, clip), Meta (llama4, llama3, llama2), Google (gemma3, gemma, t5, bert), Mistral (mistral, tekken), Alibaba (qwen), DeepSeek, Zhipu (glm), Tencent (hunyuan), Baidu (ernie), ByteDance (seed), 01.AI (yi), IBM (granite), TII (falcon3, falcon), EleutherAI (neox), BigScience (bloom), UW (superbpe), and Moonshot (kimi).
Does Tokenizer Mono work for Claude or Gemini?
No. The creator notes Anthropic's tokenizer isn't public, so there's no Claude face, and Gemini's tokenizer is also not public.
Does Tokenizer Mono require JavaScript or an internet connection?
No. All highlighting is done by the font's OpenType glyph substitution rules — no JavaScript, no network calls. It works offline in any compatible editor.
Facts on this page are quoted from george.mand.is and checked when written (9 November 2026). Sites change; check the site before you rely on details like prices.