Counting tokens for Agents-A1
InternScience ships the tokenizer with Agents-A1's Apache-licensed weights, so the counts on this page are exact. Agents-A1 (35B-A3B) and Agents-A1 4B load by default and share one count, because both ship the same tokenizer file. Both show as free: there is no hosted API, and with 3B active parameters the 35B model runs on a single machine. The training trajectories behind Agents-A1 average around 45,000 tokens, so the context bar against its 262,144-token window is the number to watch when you feed it a long tool-call history.
Pricing and context windows
Agents-A1 models with their tokenizer and context window. The weights are open under Apache 2.0 and there is no hosted API, so they show as free.
| Model | Tokenizer | Price / 1M input | Context window |
|---|
| Agents-A1 35B-A3B | agents-a1 | Free | 262K |
| Agents-A1 4B | agents-a1 | 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 Agents-A1 tokenization works
Agents-A1 is fine-tuned from Qwen3.5 and keeps the Qwen vocabulary, with the audio special tokens of the Qwen3-Omni line added. The tokenizer is a byte-level BPE with about 248,000 entries, identical in vocabulary, merges, and splitting rule to Qwen3.5's, plus seven special tokens for audio that also appear in Qwen3.8's file. Those special tokens never occur in pasted text, so Agents-A1 counts exactly like Qwen: 579 tokens on our fixed 500-word English sample. Any Qwen3.5, 3.6, or 3.8 count you already have applies to Agents-A1 unchanged.
The vocabulary was trained with a large share of code and Chinese, so tool outputs, source files, and mixed Chinese-English text tokenize compactly. Agents-A1 reasons before answering; that thinking is generated at reply time and is not part of the input count, though it consumes context.
Frequently asked questions
Is this Agents-A1's official tokenizer?
Yes. The tool loads the tokenizer file from the model's Hugging Face repository. The text is tokenized in your browser, so the count is exact and nothing is uploaded. vLLM, SGLang, and the GGUF builds use the same vocabulary, so the number matches what your own deployment reports.
Do Agents-A1 and Agents-A1 4B tokenize the same way?
Yes. The two repositories ship the same tokenizer file. The count is identical on both; they differ in size, speed, and quality, not in how text is split. The compare table shows one count against each model's context window.
Does Agents-A1 count the same as Qwen?
Yes. It keeps Qwen3.5's vocabulary and splitting rule. The only additions are special tokens that never appear in ordinary text, so a Qwen count carries over directly. The Qwen page in this tool produces the same number.
How much of the 256K context will an agent run use?
Paste the full transcript, including tool outputs, and read the context bar. Agents-A1 is built for long-horizon tasks, and the tool results in a search or coding loop usually outweigh the prompts. Leave room for the model's reasoning and its final answer, which are added on top of the input.
