Choose Instant, Fast, Smart or Genius
ZenithAI gives you tiers, not model names. You pick a tier by its name, and your server takes care of the rest. Think of it like choosing between a scooter, a car and a train: all reach the destination, at different speeds and for different loads.
| Tier | Good for | Images | Your functions (tools) |
|---|---|---|---|
| Instant | Quick replies, sorting and tagging, yes or no answers | Yes | No |
| Fast | Everyday chat, summaries, pulling fields out of text, quick agents | Yes | Yes |
| Smart | Full-quality answers, longer documents, Office files | Yes | Yes |
| Genius | The hardest work: careful reasoning over long material | Yes | Yes |
Tools on Fast come with installer 4.6.0. A server installed from an older installer takes tools on Smart
and Genius until it is updated to 4.6.0. /v1/models always shows what your server offers.
Instant, Fast and Smart are on every server. Genius is available when it is switched on for your server. Higher tiers give better answers and take longer.
Check what your server offers
Section titled “Check what your server offers”Servers can be set up differently, so ask yours before you build:
import os, requests
r = requests.get(os.environ["BASE"] + "/v1/models", headers={"Authorization": "Bearer " + os.environ["API_KEY"]})r.raise_for_status()for m in r.json()["data"]: caps = m["capabilities"] print(m["id"], "| context:", m["context_window"], "| max output:", m["max_output_tokens"], "| images:", caps["vision"], "| tools:", caps["tools"])curl --fail-with-body "https://dummydomain/v1/models" \ -H "Authorization: Bearer $API_KEY"Each tier lists:
context_window: how much text, in tokens, one request can holdmax_output_tokens: the longest answer you can ask forcapabilities:vision(images),tools(function calling),response_formatandstreaming- for image tiers,
max_images_per_request; for tool tiers,tool_context_windowandmax_output_tokens_with_tools
Context budgets
Section titled “Context budgets”The context window is how much one request can hold, counted in tokens: your question, the earlier turns of the chat, any text you attach, and the answer itself. A token is about four characters of English text. Hindi and other scripts use more tokens for the same length.
These are the defaults on a standard ZenithAI server:
| Tier | Context window | With tools active | Longest answer |
|---|---|---|---|
| Instant | 49,152 tokens (48K) | no tools | 16,384 |
| Fast | 49,152 tokens (48K) | 24,576 | 16,384, or 4,096 with tools |
| Smart | 49,152 tokens (48K) | 49,152 | 16,384, or 4,096 with tools |
| Genius | 131,072 tokens (128K) | 98,304 | 16,384, or 8,192 with tools |
Your server can differ, because your administrator can change these numbers. context_window,
tool_context_window, max_output_tokens and max_output_tokens_with_tools in
GET /v1/models are always the live values.
How far the window can be raised
Section titled “How far the window can be raised”- ZenithAI’s ceiling is 131,072 tokens (128K) on every tier. The installer and the API refuse a larger window.
- The Genius model itself is trained for 262,144 tokens, twice that ceiling. ZenithAI stops at 128K, so that extra length is not available.
- Instant, Fast and Smart share one graphics card. Raising any of them means planning for fewer people at once, or more graphics memory.
A bigger window costs graphics memory and time. See
what a larger window costs before you ask for one. Your
administrator makes the change with ZenithAI support, and /v1/models shows the new value once it is live.
What a bigger window does not change
Section titled “What a bigger window does not change”The window sets the most a request may hold. Some limits in the web app are separate and stay as they are when the window grows, such as the characters read from one attached file in a chat and the amount of document text used as evidence for one answer. If a long document is cut short, a bigger window alone may not help. Review long documents in parts, or see Document intake.
If a request is too big, the API answers context_length_exceeded
(Errors). Shorten the input, send fewer earlier turns, or move to a tier with a bigger window.
Reasoning
Section titled “Reasoning”In the web app, a Reasoning switch lets the model think before answering, on tiers that allow it. It
is off by default. On the API, Genius thinks first when a request sends reasoning_effort (Chat Completions)
or reasoning.effort (Responses) as low, medium, high or xhigh; otherwise the API answers directly.
See thinking on Genius.
Common questions
What is the difference between Fast, Smart and Genius?
Fast gives balanced, quick answers for everyday work. Smart gives full-quality answers and can call your functions. Genius is the top tier for the hardest work, such as long documents and careful reasoning. Higher tiers are slower.
Which ZenithAI tiers can read images?
By default every tier accepts images, up to 8 per request and 20 MiB in total. Your administrator can change this, so check the vision capability in GET /v1/models.
Which tiers support function calling?
Fast, Smart and Genius take your functions on /v1/chat/completions and /v1/responses. Instant is kept for quick replies. GET /v1/models shows tools as true for every tier on your server that supports them.
How big is the context window of each ZenithAI tier, and can it be raised?
On a standard server Instant, Fast and Smart take 49,152 tokens and Genius 131,072. ZenithAI allows at most 131,072 tokens on any tier, although the Genius model is trained for 262,144. Your administrator can change the window with ZenithAI support; GET /v1/models shows the live value.
