Think of it as reading speed
When MANU answers a question, text appears in the chat one piece at a time. Token rate (tok/s) measures how many of those pieces show up each second.
The easiest way to judge whether a number is “fast enough” is to compare it to how quickly you can comfortably read technical text — not to raw computer throughput.
How we estimate comfortable reading speed
- ~200 words per minute for dense technical prose — a middle estimate; research often cites roughly 150-250 WPM for non-fiction.
- ~1.3 tokens per word for English (about 0.75 words per token) — a standard LLM approximation.
200 WPM ÷ 60 ≈ 3.3 words/sec x 1.3 tokens/word ≈ roughly 3-5 tok/s for comfortable reading. Skimming is faster; carefully studying specs, tables, or diagrams is slower.
If MANU streams at 17 tokens per second, the answer appears faster than you can read it. You won't be waiting on the text to catch up.
| Speed | What it feels like |
|---|---|
| ~4 tok/s | Comfortable reading of technical text |
| 17 tok/s (Desktop) | About 4× faster than you can read — answers keep ahead of you |
| 100 tok/s (Team) | Far above reading speed; useful when multiple people chat at once |
| 150 tok/s (Server) | Highest headroom for teams and heavier workloads |
What is a token?
A token is a small chunk of text — often part of a word, a whole word, or a short phrase. Language models process and generate text in tokens rather than whole sentences at once.
Tokens per second means how many of those chunks appear in the answer stream each second. See MANU pricing for tok/s by appliance tier.
Behind the scenes
For readers who want the full picture — including work you don't see in the chat.
Technical details
Not all tokens are visible
Before and during an answer, the model may use tokens for retrieval, assembling context from your manuals, and internal reasoning. That work does not show up in the chat stream you read.
"Thinking" and time to first word
Some models reason before producing user-visible output. That can mean a short pause before text starts streaming — even when tok/s is high once streaming begins. Token rate measures streaming speed, not total response time from when you press send to when the last word appears.
Indexing vs answers
MANU reports two different speeds on pricing:
Indexing tok/s — how fast MANU processes documents when you add them (up to 10,000 / 100,000 / 200,000 tok/s by tier).
Answers tok/s — how fast chat replies stream (up to 17 / 100 / 150 tok/s by tier).
"Up to" disclaimer
Listed rates are hardware-tier maximums. Actual speed varies with question complexity, document size, and concurrent use.
Token rate questions
What is a good token rate for chat?
For most people, anything above roughly 3-5 tokens per second is faster than comfortable reading of technical text. MANU Desktop streams at up to 17 tok/s, so the answer appears well ahead of how fast you can read it.
Why is there a pause before the answer starts?
Before visible text appears, MANU may retrieve relevant passages from your manuals and the model may reason internally. That work does not show in the chat stream. Token rate measures how fast text appears once streaming starts, not total response time.
Is Desktop's 17 tok/s enough?
Yes, for a single user. At 17 tokens per second, answers stream roughly four times faster than typical technical reading speed. Higher tiers add headroom for multiple simultaneous users or heavier workloads.
What's the difference between indexing and answer speed?
Indexing tok/s is how fast MANU processes documents when you add them to the library. Answer tok/s is how fast chat replies stream to your screen. Both are listed on the MANU pricing page by appliance tier.
See how MANU tiers compare
Desktop, Team, and Server — answer speed, storage, and more.