Counting tokens for Ornith models
Ornith ships its tokenizer with the MIT-licensed weights, so the counts on this page are exact. Ornith 1.5 9B and 35B-A3B load by default. Add the 397B model or any Ornith 1.0 size from the compare table to count the same text on them. Every Ornith model shows as free: there is no first-party API, the weights are open, and the 9B and 35B sizes run on your own hardware. The figure that matters is the context bar against the 262,144-token window. Agent transcripts grow fast, and the bar tells you how much room a repository dump or a long tool-call history leaves.
Pricing and context windows
Ornith models with their tokenizer and context window. There is no first-party hosted API and the weights are open under MIT, so they show as free.
| Model | Tokenizer | Price / 1M input | Context window |
|---|
| Ornith 1.5 9B | ornith-1.5 | Free | 262K |
| Ornith 1.5 35B-A3B | ornith-1.5 | Free | 262K |
| Ornith 1.5 397B | ornith-1.5 | Free | 262K |
| Ornith 1.0 9B | ornith-1.0 | Free | 262K |
| Ornith 1.0 35B | ornith-1.0 | Free | 262K |
| Ornith 1.0 397B | ornith-1.0 | Free | 262K |
Prices and context windows are curated metadata refreshed periodically from the providers and OpenRouter, so treat them as guidance rather than a quote. Token counts are always produced by the real tokenizer.
How Ornith tokenization works
Ornith is built on Qwen3.5, and Ornith 1.5 ships the Qwen3.5 tokenizer file unchanged. Ornith 1.5 (9B, 35B-A3B, and 397B) is post-trained from Qwen3.5 with additional pretraining, and its tokenizer is byte-identical to the file Qwen publishes: a byte-level BPE with about 248,000 entries. Ornith 1.0 uses the same vocabulary and merge rules, but its file carries seven extra special tokens for audio, inherited from the Qwen3-Omni line, and an older word-splitting rule that does not treat combining marks as part of a word. On plain text the two versions count the same: 579 tokens on our fixed 500-word English sample, matching Qwen. They differ on scripts that use combining marks: on a Hindi or Thai sentence, Ornith 1.0 needs close to twice as many tokens as 1.5, which is why this page keeps the two versions as separate tokenizers.
The Qwen vocabulary was trained with a large share of code and Chinese, so source files and mixed Chinese-English text tokenize compactly. Ornith answers with a thinking block by default; that reasoning is generated at reply time and is not part of the input count shown here, but it does consume context.
Frequently asked questions
Do Ornith 1.0 and 1.5 use the same tokenizer?
Almost. The vocabulary is the same; 1.0 adds audio special tokens and uses an older splitting rule. For English, code, Chinese, and precomposed scripts such as Vietnamese the counts match. Scripts built from combining marks, such as Devanagari or Thai, split into far more tokens on 1.0. Add both versions from the compare table to check your own text; if the numbers match, either count is safe to use.
Does Ornith count the same as Qwen?
Yes. Ornith 1.5 ships Qwen3.5's tokenizer file byte for byte. Any Qwen3.5 or Qwen3.6 count applies to Ornith 1.5 directly, and the Qwen page in this tool gives the same number. Ornith 1.0 matches on ordinary text as well.
How much of Ornith's 256K context will my text use?
The context bar above the editor shows the share of the 262,144-token window your text fills. Ornith is meant for long agentic coding sessions, so the input is usually a repository, a diff, and a tool-call history rather than a single prompt. Paste the transcript here to see how much room is left for the model's thinking block and its answer.
Which Ornith models does this cover?
Ornith 1.5 9B, 35B-A3B, and 397B, plus Ornith 1.0 9B, 35B, and 397B. Sizes within a version share one tokenizer, so the count also applies to the quantized GGUF, MLX, and FP8 builds of each model.
