Repository navigation
Replies: 2 comments 1 reply
|
@cmuentes Hi, this looks like a bug. Also, could you please reset the generation settings to their defaults (ellipsis in the top-right corner → Generation Settings → Reset / Reset to System Settings) and see if that helps? We have a Poco F8 Ultra device as well, so I will take a look and share the resutls. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment


Uh oh!
There was an error while loading. Please reload this page.
Hello - I'm trying to set up a Perplexity-alternative using my Poco F8 Ultra with 16gb of RAM. I have the Gemma 4 E2B and E4B models working but for both models, but especially for E4B, my phone gets extremely warm (not too hot to the touch or anything) and I can watch my battery drain even though this phone has a pretty large battery. For that reason I've been looking for other small models, especially 4B models, that can maybe work as a "Flash" model for quick answers without the excessive heat or battery draining.
So I decided to try some of the models the PocketPal AI team wrote a blog post on, featuring models that did a good job with tool invocation and factual responses, but when I try them in the app, I get junk responses (see screenshots). I tried with a system prompt in a Pal, and without a Pal, and as you can see the models, specifically LFM2.5 1.2B Instruct, Qwen3.5 4B, and LFM2.5-2.6B but with and without Thinking on, each model's reasoning is broken. If Thinking is off, the model just outputs the garbage response as a chat response. What could I be missing? Again the Gemma 4 models work but burn through battery so I'm wondering what else I can do.
All reactions