# Ask one question. Watch two builds answer it.

A live, side-by-side comparison of two builds of the same model: Tileward Qwen3.8-27B in TWFP4 and NVIDIA's Qwen3.8-27B in NVFP4. Both answer the same question, from the same documents, at the same moment, and the speed and token figures are measured live on each run.

Page: https://tileward.com/use-cases/compare/

## What you can try

- **Pick a use case.** Each one grounds both models in its own document set:
  - Financial services: a bank's internal policy and procedure documents.
  - Healthcare: FDA prescribing information, standing in for a hospital formulary.
  - Customer support: airline, retail and telecom support handbooks.
  - Public sector caseworker: Social Security, VA, DMV and Federal Student Aid pages.
  - Semiconductor: US export control rules and a fab's defect SOPs.
- **Ask a question.** Choose one of the sample questions or ask your own about those documents, then ask both models at once. A sample question marked with a lock touches a tile locked for that use case.
- **Switch Reasoning on** to have both models reason before they answer; each side then also reports its reasoning time.
- **Read the result.** Each side shows its answer and its own figures, and the page lists the sources given to both sides.

The comparison runs in the browser with JavaScript on.

## What each side reports

- Time to first token
- Reasoning time
- Time to last token
- Tokens per second (decode)
- Tokens (prompt / completion / total)

## How the numbers are measured

Both sides run on one clock. Each question is checked once and grounded once. Both models then receive the same messages and settings at the same moment, and timing starts at the gateway, so your own network sits outside every figure.

| Figure | What it measures |
|---|---|
| Time to first token | From the request leaving the gateway to the first token back, reasoning or answer. |
| Reasoning time | From the first reasoning token to the first answer token, when Reasoning is on. |
| Time to last token | From the request leaving the gateway to the last token back. |
| Tokens | The engine's own count for the prompt and the completion. The completion includes any reasoning. |
| Tokens per second | Completion tokens after the first, divided by the time from first token to last: the speed of generation. |

## On your own documents

Run the same test on the questions your team asks: bring the documents and the questions, and an engineer runs them with you and shows you the numbers. Thirty minutes, no deck. [Talk to an engineer](https://calendar.app.google/hg2DzFTxKJgwFWm19).

## Related

- [Use cases](https://tileward.com/use-cases/): the three job patterns and worked examples
- [Tileward Models](https://tileward.com/models/): run larger models on existing GPUs
- [Model library](https://tileward.com/model-library/): every model, its size and its price
- [Start free](https://app.tileward.com/?join=explore): the Tileward console, no card