Model download might be unstable sometimes

[Amio Jin]

After initiating the model download, the progress events popped up and then stopped. For a while, it throws "loaded: 1". See the screenshot.

I found that the download progress is often unstable, and it takes some time; users may leave the page during the download. So there's a significant chance that when (user reopen the page) we call LanguageModel.availability(), it's "downloading". In this case, how can we check the download progress?

[François Beaufort]

On Thu, Sep 11, 2025 at 11:27 AM Amio Jin <ami...@gmail.com> wrote:

After initiating the model download, the progress events popped up and then stopped. For a while, it throws "loaded: 1". See the screenshot.

As you can read in https://chromium-review.googlesource.com/c/chromium/src/+/6787415, the download progress of the model has changed to only account for 90% of the progress bar.
The remaining 10% (may be more if the model is partially downloaded) won't load until the create() call resolves successfully.
That's because when the model was finished downloading (e.g. loaded = 1.0), there was still plenty of time before the call to create() successfully resolved asynchronously, leaving the user with an unknown expectation of waiting time. This was due to the uncompression of the model on the device. 

We're currently working on improving the download progress story as this process can take a looooonnng time.
 

I found that the download progress is often unstable, and it takes some time; users may leave the page during the download. So there's a significant chance that when (user reopen the page) we call LanguageModel.availability(), it's "downloading". In this case, how can we check the download progress?

You can call create() again with a monitor option to get downloadprogress events again.
 

--
You received this message because you are subscribed to the Google Groups "Chrome Built-in AI Early Preview Program Discussions" group.
To unsubscribe from this group and stop receiving emails from it, send an email to chrome-ai-dev-previe...@chromium.org.
To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/chrome-ai-dev-preview-discuss/48136f4a-f17b-4c0b-93a3-a5c8e55d979en%40chromium.org.

[Amio Jin]

Thanks François!

In the screenshot it stops at 0.04, a lot less than 90%. Might be a bug or something?

But it's good to know we can call create() again for the remaining events, at least this is recoverable.

[François Beaufort]

On Fri, Sep 12, 2025 at 2:11 AM Amio Jin <ami...@gmail.com> wrote:

Thanks François!

In the screenshot it stops at 0.04, a lot less than 90%. Might be a bug or something?

It definitely sounds like a bug.
Can you file one at http://crbug.com/new?component=1583300 so that the Chrome engineering team is aware of those behaviors?

[Amio Jin]

Sure, filed it here https://issues.chromium.org/issues/444560478

[François Beaufort]

Thank you! I've triaged the bug, hopefully we'll hear from the team soon.