🚀Pocket LLM v1.5.0 is out: offline Android LLM chat with voice, image input, OCR, and camera capture

Posted by 100daggers_@reddit | LocalLLaMA | View on Reddit | 24 comments

Pocket LLM v1.5.0🚀

New in this release:

- 🎙️ Voice input

- 🖼️ Image input with OCR, Gemma vision, and FastVLM support

- 📷 Camera capture with retake, crop, and photo review

- 🗂️ Previous chats side panel

- 💾 Downloaded model deletion to save storage

- ⚙️ Editable model instructions with presets and custom prompts

- 🎨 Light/dark mode, accent colors, and font-size controls

- 📋 Copy option for assistant responses

[-]

f4nt4@reddit

Looks great and runs smoothly!

Do you have any plans to provide an OpenAI compatible endpoint, to have your app running as a server? Also a tk/s statistic would be pretty nice!

[-]

RickyRickC137@reddit

It will be great if we can use our own models instead of mandatory downloading at the start up. Kinda annoying.

[-]

WolfeheartGames@reddit

https://play.google.com/store/apps/details?id=com.pocketclaws.app bring your own model. Significantly more feature rich than anything else.

[-]

100daggers_@reddit (OP)

Fair point, I’ll keep this in mind for the next update.

When you say “use our own models,” do you mean custom fine-tuned models, or any compatible model from Hugging Face?

I avoided open-ended model loading for now because it adds a lot of complexity: formats, tokenizers, prompt templates, memory limits, and backend differences. My current goal is to keep the app simple for users who just want a working offline model without tuning tokens, temperature, instructions, or compatibility settings.

For advanced users, custom model support could make sense later, but I want the default experience to stay reliable and easy.

Also curious what others think: do most people want custom model support, or mainly a simple offline app that just works?

[-]

RickyRickC137@reddit

I understand your situation. I can only speak for myself. I think I would like to add custom models because I already have pocketpal and I downloaded heretic models (because Gemma e2/4b models are super censored and defeats the purpose of running on mobile). I don't wanna download models again for a new app. Plus there are few key parameters that even simpletons like myself would like to tweak, such as, temp, repeat and presence penalty, min p, top k, (as we've seen in LMstudio) and I don't think they're pretty advanced (you can set them at default value to help newbies). I like your GUI. And if you can add features like STT (whisper), TTS (kokoro, kitten), web search (with duck duck go, brave, searxng), it would gain a lot of traction.

[-]

100daggers_@reddit (OP)

Thanks for clarifying. I’ve added this to my list.

Especially support for pointing to an already downloaded custom model, model statistics, and generation/model settings.

[-]

RickyRickC137@reddit

Awesome. Would love to beta test it if needed!

[-]

rorowhat@reddit

Nice!

[-]

KaroYadgar@reddit

looks kinda like AI-generated slop but hey, it's open-source so I can't complain. Looks like a good product, so good on you!

UI reminds me a lot of an older version of my own site, especially the layout, color scheme, icons, and the terrible-to-look at chat deletion icon to the right of every conversation title (which I've thankfully fixed, you should too.)

[-]

100daggers_@reddit (OP)

Thanks for the honest part of the feedback. I agree the chat deletion icon can be improved, and I’ll clean that up in the next UI iteration.

But calling it “AI-generated slop” is unfair. I’ve been working on this app for around a year. AI has helped me move faster, like it helps many developers now, but the product direction, architecture, debugging, model integration, Android implementation, and release work are not just some generated template.

Using AI-assisted coding in 2026 is normal. Dismissing someone’s work as “AI slop” without knowing the effort behind it does not add much value.

Specific UI feedback is appreciated. The comparison to your own site and the “can’t complain” framing felt more like a jab than useful feedback.

[-]

KaroYadgar@reddit

I was too harsh. I intended the original message to be a joke but I completely understand how it could be interperated as an insult. It's clear you put a lot of effort into it and it definitely looks like a great product. The frontend reminded me of older model frontends, which is what led me to making the comment.

I've also been working on my UI (on and off) for about a year too, that's awesome. Mine has gone through many phases with terrible looking UI, some looked ridiculously ugly. I wish you the very best, and I'll probably try it out myself.

[-]

100daggers_@reddit (OP)

No worries, I appreciate you clarifying.

Thank you, and please let me know if you have any other feedback. All the best to you too.

[-]

Dhonnan@reddit

Wheres the link where i can download this?

[-]

100daggers_@reddit (OP)

https://github.com/dineshsoudagar/local-llms-on-android/releases/tag/v1.5.0

[-]

MalabaristaEnFuego@reddit

What percentage of code does a person need to refactor before the code no longer becomes AI slop?

[-]

handsoapdispenser@reddit

Hank Green actually gave a really solid breakdown of what is and isn't slop. It really boils down to a combination of human effort, intent and utility.

[-]

MalabaristaEnFuego@reddit

I thought it was very interesting that he emphasized how there was significantly less backlash against the coding community, but I think he also somewhat glossed over the fact about how structured most code is in the most practical sense. I've tested the same coding prompt across 20 different models of various sizes multiple times for performance testing and calibration, and the variance between the outputs is essentially a rounding error at this point.

I feel like coders, and people who are starting to use AI for coding workflows that may be new to coding, are using more sophisticated workflows than the average "AI Slop" generator. I have a working catalog of who my strongest local LLMs are for coding in their respective areas. I build plans with smarter models co-creating the plans with me and delegating the work to open source models, then passing back through me and the smart model before handing it off to the flagship coder for final polish. I still review all code line by line every step of the way, run tests in sandboxes, etc.

[-]

handsoapdispenser@reddit

I think that goes exactly to his point though. There are creators posting videos to tiktok for likes that are absolutely using AI to generate every frame but they are also definitely calibrating their prompts and testing outputs. Not to credit evil propagandists but those Iranian Lego videos are definitely not slop.

[-]

KaroYadgar@reddit

slop of Theseus

jokes aside, I was mainly referring to how the UI reminded me of AI-generated frontends back in the day.

[-]

MalabaristaEnFuego@reddit

lol, Slop of Theseus. Nice.

I'm trying to build an inventory management system from scratch and the GUI is definitely turning out to be the difficult bit that I need to be most manual and hands on with, and it's primarily more intended to be functional over form at this point.

Building the back end felt like a breeze compared to the front end.

[-]

KaroYadgar@reddit

Hah, I guess that feeling is common for all of us. AI certainly helps with frontend, as long as you know how to steer it well enough.

[-]

handsoapdispenser@reddit

App store or GitHub link?

[-]

100daggers_@reddit (OP)

https://github.com/dineshsoudagar/local-llms-on-android/releases/tag/v1.5.0

[-]

Effective-Drawer9152@reddit

Gemma 4 E2B is pretty good for my 8GB phone