prefix behavior has changed?

[Die4Ever2005]

let m = await LanguageModel.create();

await m.prompt([{role:'user', content: 'Hello'},{role:'assistant', content: 'Hello', prefix:true}]);

"Hi there! 👋  Glad to hear from you.  Is there anything specific you'd like to talk about or need help with? I'm happy to answer questions, generate ideas, write stories, or just chat.  \n\nLet me know! 😊\n\n\n\n"

If I prepend the prefix text, which is what I used to do, then it would look like "HelloHi there!"

[Die4Ever2005]

Version 144.0.7559.31 (Official Build) beta (64-bit)

Foundational model state: Ready
Model Name: v3Nano
Version: 2025.08.14.1358
Backend Type: CPU
File path: /home/ray/.config/google-chrome-beta/OptGuideOnDeviceModel/2025.8.21.1028
Folder size: 4,236.82 MiB

[Die4Ever2005]

I tried it again and similar result.

let m=await LanguageModel.create(); await m.prompt([{role:'user', content: 'Hello'},{role:'assistant', content: 'Hello', prefix:true}]);

"Hi there! 😊 \n\nHow are you doing today?  Is there anything I can help you with?  \n\nI'm ready for anything you might ask or discuss!  Whether you need information, creative writing, help brainstorming, or just a friendly chat, I'm here. 😊\n\n\n\n"

chrome://on-device-internals/ says:

2:34:47 PM ai_language_model.cc(544) Starting on-device session for PromptApi
2:34:48 PM ai_language_model.cc(254) Executing model with input context of 14 tokens: <user>Hello<end><model>Hello
2:34:56 PM ai_language_model.cc(364) Model generates raw response with PromptApi: Hi there! 😊 How are you doing today? Is there anything I can help you with? I'm ready for anything you might ask or discuss! Whether you need information, creative writing, help brainstorming, or just a friendly chat, I'm here. 😊

[Die4Ever2005]

Interesting, it works fine on my desktop computer.

let m = await LanguageModel.create(); await m.prompt([{role:'user',content:'Hello'},{role:'assistant',content:'Hello',prefix:true}]);
23:02:28.582 " there! 👋\n\nHow can I help you today? I'm ready to answer your questions, chat, brainstorm ideas, or anything else I can do. 😊 \n\nJust let me know what's on your mind!"
23:02:41.360 let m = await LanguageModel.create(); await m.prompt([{role:'user',content:'Hello'},{role:'assistant',content:'Hello',prefix:true}]);
23:02:43.075 " there! 👋\n\nIt's nice to hear from you. How can I help you today?  Do you have any questions, need some information, or just want to chat? 😊 \n\nI'm ready for anything! Just let me know.\n\n\n\n"
23:02:49.281 let m = await LanguageModel.create(); await m.prompt([{role:'user',content:'Hello'},{role:'assistant',content:'Hello',prefix:true}]);
23:02:51.083 " there! 👋 \n\nIt's nice to hear from you. How can I help you today? 😊 \n\nI'm ready to answer questions, generate text, translate languages, write stories, or just chat. Just let me know what you have in mind! \n\n"

Version 144.0.7559.20 (Official Build) beta (64-bit)

Foundational model state: Ready
Model Name: v3Nano

Version: 2025.06.30.1229
Backend Type: GPU (highest quality)
File path: /home/ray/.config/google-chrome-beta/OptGuideOnDeviceModel/2025.8.8.1141
Folder size: 4,088.14 MiB

[Die4Ever2005]

My desktop is using an older model, OptGuideOnDeviceModel/2025.8.8.1141

Maybe it failed to update due to the Linux issue with the size of /tmp and I haven't been starting Chrome with the custom TMPDIR envvar.

[Sumeet Singh]

Hi Die4Ever2005,

Thank you for providing the additional details.

It seems the different prefix behavior might be related to the model version you are running:

  - The older model version (2025.08.08.1141) on your desktop works as expected.

  - The newer model version (2025.08.21.1028) on the other machine exhibits the changed behavior.

This suggests a change in how the newer version of the model processes the prefix: true attribute.

Best regards,

Sumeet pal singh 

--
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/28ab8ae4-4060-4470-b4c9-94744fefd164n%40chromium.org.

[Die4Ever2005]

Yes it seems to be a regression, a new bug.

I'm refraining from updates for now (by utilizing the bug where Chrome is unable to update the model on Linux when your /tmp partition is too small lol)

Hopefully someone posts here when it's been fixed so I know when it's safe to update again.

[Thomas Steiner]

I have created a spec issue, as, to me, it doesn't immediately look clear what the expected behavior is (not from the explainer nor from the spec draft). 

To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/chrome-ai-dev-preview-discuss/585bface-5519-4654-af1e-4b5f3bc33929n%40chromium.org.

--

Thomas Steiner, PhD—Developer Relations Engineer (blog.tomayac.com, toot.cafe/@tomayac)

Google Spain, S.L.U.

Torre Picasso, Pl. Pablo Ruiz Picasso, 1, Tetuán, 28020 Madrid, Spain

CIF: B63272603

Inscrita en el Registro Mercantil de Madrid, sección 8, Hoja M­-435397 Tomo 24227 Folio 25

----- BEGIN PGP SIGNATURE -----

Version: GnuPG v2.4.8 (GNU/Linux)

iFy0uwAntT0bE3xtRa5AfeCheCkthAtTh3reSabiGbl0ck

0fjumBl3DCharaCTersAttH3b0ttom.xKcd.cOm/1181.

----- END PGP SIGNATURE -----