> ## Documentation Index
> Fetch the complete documentation index at: https://modal-computer-use.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Weighted input-capacity result

> Review the evidence behind the 100-token refill and 400-token burst defaults.

## Use the promoted default

Modal Computer Use 2.0.1 uses a refill of 100 normalized input-work tokens per second and a
400-token burst. The daemon reserves the complete recursive cost of an ordered batch before any
desktop mutation. A transient rate limit therefore cannot interrupt an admitted batch halfway
through.

The `normalized-input-work-v1` policy assigns these costs:

| Input action | Tokens |
| - | -: |
| Move, mouse down or up, keypress, release all | `1` |
| Click, double click, triple click | Click count |
| Hotkey | `max(1, ceil(keys / 4))` |
| Type | `1 + ceil(characters / 32)` |
| Scroll | `1 + ceil(amount / 32)` |
| Coordinate drag | `1` |
| Path drag | `1 + ceil(path points / 32)` |
| Hold key with nested actions | `1 + sum(nested action costs)` |
| Wait, screenshot, zoom, cursor query | `0` |

These weights estimate input work for resource admission. They do not replace application approval
or safety policy.

## Review the three passing runs

The minimum tested Sandbox used one CPU and 2,048 MiB. Each independent run completed 80 batches
of 48 ordered mixed input actions. Each batch cost 60 normalized tokens.

| Run | Weighted tokens/sec | Batch p50 | Batch p95 | CPU seconds/token | RSS growth | Result |
| - | -: | -: | -: | -: | -: | - |
| 1 | 527.398 | 113.414 ms | 131.890 ms | 0.007156250 | 0 B | Pass |
| 2 | 505.135 | 116.777 ms | 148.103 ms | 0.007408333 | 3,817,472 B | Pass |
| 3 | 380.704 | 149.609 ms | 206.166 ms | 0.008381250 | 29,941,760 B | Pass |

All 240 measured batches returned 48 ordered successful results with native XTest attribution and
the expected pointer sentinel. The artifacts contain no failures, retries, replacement samples, or
cleanup survivors. The slowest run sustained 380.704 tokens per second.

## Keep the gate separate from the default

The benchmark configured a diagnostic 2,000-token refill and 4,000-token burst. That limiter stayed
outside the measured backend capacity. It is not the product default.

Each run had to sustain at least 200 normalized tokens per second. It also had to use no more than
0.02 aggregate cgroup CPU-seconds per normalized token and add no more than 128 MiB of RSS. Exact
placement, resources, ingress, XTest attribution, ordered results, and clean lease release were
required. The measured source revision was
`eee2b9456c76474a5b50a857af899ff11ca70a32`.

Do not raise the 100/400 product values from this result alone. Run the same-runtime capacity gate
for your action mix, browser load, resources, and placement.

## Inspect the release evidence

* [Dated input-capacity report](https://github.com/ashtonchew/modal-computer-use/blob/b60c1cb7495200e36a738c0f6e07961b1d2db93c/docs/benchmark-results-2026-08-08-input-capacity.md)
* [Passing run 1](https://github.com/ashtonchew/modal-computer-use/blob/b60c1cb7495200e36a738c0f6e07961b1d2db93c/benchmark-data/input-capacity-run-1-2026-08-08.json)
* [Passing run 2](https://github.com/ashtonchew/modal-computer-use/blob/b60c1cb7495200e36a738c0f6e07961b1d2db93c/benchmark-data/input-capacity-run-2-2026-08-08.json)
* [Passing run 3](https://github.com/ashtonchew/modal-computer-use/blob/b60c1cb7495200e36a738c0f6e07961b1d2db93c/benchmark-data/input-capacity-run-3-2026-08-08.json)

The artifacts are sanitized and immutable. The application repository owns the benchmark runner,
validation rules, and source evidence.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.