Best world knowledge model that can run on your phone

[-]

fraterdidymus@reddit

There isn't and won't be one, because LLMs do not contain knowledge. The only way you can ever tell the difference between a fact and a hallucination in an LLM output is by already knowing the answer yourself.

I'm sorry, but LLMs are simply not "a super efficient compression algorithm for the entire internet".

[-]

deject3d@reddit

I thought your take was pretty dumb so I asked chatGPT's o3 model to defend itself from your points:

-- AI response below:

1) “LLMs don’t contain knowledge.”

If that were literally true, closed-book QA benchmarks (where the model gets no web access or retrieval) would be impossible. Yet GPT-4, Claude 3, etc. routinely hit 70-90 % on datasets like Natural Questions and TriviaQA—numbers that beat most humans⁠.¹ A model can’t answer “What country did the Dreyfus Affair take place in?” or “Who coined the term ‘meme’?” without some internal representation of those facts, however fuzzy.

2) “You need to know the answer already to spot hallucinations.”

That’s true for any information source. If you don’t already know, you:

cross-check (Wikipedia, a primary source, a textbook);
triangulate (ask a second expert—or, yes, a second model);
look for citations (which newer models can emit automatically).

Humans hallucinate constantly—see eyewitness-testimony research—but we still treat conversation as informative because we have external validation loops. The same hygiene works for LLMs.

3) “Not a super-efficient compression of the internet.”

It depends on what you mean by “compression.” Training does squeeze hundreds of terabytes of text into a few dozen gigabytes of weights. That’s lossy (like JPEG, not ZIP), yet highly structured: common patterns survive, rare noise is dropped. Information theorists literally frame this as minimum-description-length compression,² because predicting the next token well entails capturing the underlying distribution concisely.

No one claims the weights are a bit-perfect archive, but calling the process unrelated to compression ignores a huge swath of ML literature.

[-]

tkenben@reddit

That's if you take everything that was said literally. The statement "LLMs don't contain knowledge" was not a literal statement. Everything contains some amount of knowledge. So if you ask an LLM to contradict that statement it will rely on every piece of training data it has where " don't contain knowledge" and what was said about that, and be able to argue its case without using logic. Same thing with point 2. The post didn't say that *only* LLMs you must check the answer with; rather that it is more likely a requirement compared to other resources. You've demonstrated nothing interesting here. What the person was saying made perfect sense to me. I can see where a nitpicker - in this case, the chat engine - would be able to easily dispense with nuance and nitpick.

[-]

fraterdidymus@reddit

So you thought my take was dumb, but you weren't smart enough to articulate your objections yourself? That's precious.

[-]

deject3d@reddit

Actually, I just didn't want to waste my time on you

[-]

pixelizedgaming@reddit

did your wife's boyfriend write this response for you too

[-]

Inevitable_Mistake32@reddit

I think you're asking a bit much to be honest.

We're barely at the advent of llms on our phones that are even worth having at all. Now you want internet-level knowledge in your pocket while cooking marshmallows.

The desire is great.

Realistically, download a library from kiwix, download pdfs/ebooks, and just save a repo to a folder on your phone. FAR FAR less likelyhood of being gaslit by an LLM that has no actual knowledge built in.

[-]

Anothertech4@reddit

Realistically, download a library from kiwix, download pdfs/ebooks, and just save a repo to a folder on your phone. FAR FAR less likelyhood of being gaslit by an LLM that has no actual knowledge built in.

Just so I understand this correctly, have a service scan the folder for solutions based on the documents/pdfs on the folder and allow it to summaries solutions from the information it found?

Is this the correct way of preventing AI models from getting info from unvetted sources? I wonder if this can be done with the university's library database.

[-]

ShengrenR@reddit

If you're new to this, 'the AI' isn't doing any of the finding/getting of information - this is RAG in its general form: you have to build the search/match/extract application that can get that information out of sources for you. GraphRAG, simple vanilla vector likeness, 'deep research' - you name it, there's a million flavors, but you need to build the connector to that library database and have an intelligent way of filtering. If you don't mind waiting on answers, I'd recommend looking into local deep research patterns; a quick google pulled up https://github.com/LearningCircuit/local-deep-research (I've not used it) as potentially a decent starting point.. extend that to interact/connect to your local Uni's API (go bug the librarians, they love that stuff) and off you go.

[-]

Inevitable_Mistake32@reddit

kiwix is offline wiki. you just download the datasets you want, like field medicine, survival wikis, nav maps, etc.

You CAN make your LLMs reference your wiki, but that is either some really RAG model or some kind of llamaindex creation or somewhere inbetween. Though this isn't an OOTB solution and likely won't run well on a phone device if at all.

[-]

Character_Dirt851@reddit

We're barely at the advent of llms on our phones that are even worth having at all. Now you want internet-level knowledge in your pocket while cooking marshmallows.

I mean a modern phone can fit a lot of knowledge. You can easily fit the entire Wikipedia in one, even with all media.

[-]

MLDataScientist@reddit

There is MobiRAG. If you have everything you need in multiple PDFs, it will scan for relevant parts and gives you the answer. I have not tested it personally but that is the way to do it offline - https://github.com/nishchaljs/MobiRAG

[-]

darkpigvirus@reddit

Gemma 3n you can download the android apk in github

[-]

usernameplshere@reddit

I got a phone with 16GB of RAM and there are 24GB models out there. Running any recent 7/8B model should do the trick. Just give it a try. Personally, I'm using gemma 3 4b for very quick responses on the go.

[-]

fdg_avid@reddit

Sounds like you want a book on clams.

[-]

Inside_Mind1111@reddit

Okay, identifying different clam species involves examining their anatomy. Here's a guide to the key features to look for:

Shell Shape & Structure

General Shape: Is it elongated and streamlined (like a typical steamer clam or Quahog)? Or is it broader and flatter (like some Tridacnid Clams)? Shape (Specific): Pear-shaped (e.g., Mercenaria)? Cock's foot shaped (e.g., some Ostreidae)? Tridacnid clams have distinct mantle cavity structures but their shells are often large and flat or fan-shaped. Thickness (Wall): Is the shell thick-skinned (e.g., many Oysters)? Or relatively thin-shelled (e.g., many Clams like Macella)? Is it heavy-duty or light? Surface Texture: Is the periostrium (earthy, hairy coating) present? What is the general texture? Smooth? Ribbed? Grooved? Keeled? With growth lines? Edge: Is it abruptly edged (Acute margin)? Or rounded (Convex margin)?

Hinge

Type: Is there no hinge (Clavata)? Is there a simple hinge (Plicitraria)? Is there a complex hinge (Arcuatula)? Ligament: Is it strong and wide (e.g., some Clavata)? Is it narrow (e.g., Ostrea)? Beaks: Are the anterior and posterior adductor muscles' attachment points (beaks) distinctively shaped or sized? Adductor Muscle: Is it curved (simple)? Or straight (complex)?

Mantle Edge

Shape: Is it broadly rounded (e.g., Macella)? Or narrowly pointed (e.g., Mercenaria)?

blinks

Shell Color & Pattern

Color: Is it white? Pink? Cream? Brown? Orange? Blackish (from periostrium)? Mottled? Pattern: Are there spots? Bands? Growth lines?

Interior (Umbo & Mantle Edge)

Color: Is it white? Pink? Creamy? Yellowish? Rust/orange? Purplish? Often the interior color can be a key identifier (e.g., Tridacna crocea has bright interior color patterns). Texture: Is it smooth? Ribbed? Keel-edged?

How to Use These Features

Combine observations: Look at the shell shape AND hinge type. Consider the color AND interior color. Look at the overall impression: Is it a typical clam shape with a simple hinge? Or does it look like an oyster or a Tridacnid?

Important Note

These descriptions are simplified. Many species have subtle variations within their general category. For confident identification, especially of rare or specific species, you might need:

Access to detailed field guides or identification keys. Comparing photographs or actual specimens. Consulting experts or museum collections.

Handle clam shells carefully, especially if from the wild or potentially contaminated sources.

[-]

vtkayaker@reddit

The preview of Gemma 3n 4B (effective) runs on my phone, and it performs more like a solid 8-12B in my preliminary testing. (There's some really funky quantization and other stuff going on.)

It can describe photos, do light OCR, translate from French to English, write very mediocre fictional scenes, etc.

To run it, you need a special Android app that is installed as an APK. And there's no history of previous chats. Very bare bones UI for now.

[-]

Hyiazakite@reddit

This is not how LLM:s work. LLMs are very inefficient in presenting information and with large amount of parameters it requires large amount of compute. It's also not a compression algorithm. If you want to have access to a lot of information offline download Kiwix

[-]

ArsNeph@reddit

Unfortunately, no small model will be able to accurately answer those questions without massive hallucination. That said, in that size range, Gemma 3 4 is the best in terms of world knowledge, and also multimodal. Qwen 3 4B is a reasoning model and not multimodal, as well as weak in world knowledge, so probably not what you're looking for.

I think the best approach to what you want to do is to use Gemma 3 4B and ground it using RAG. You can download a complete offline copy of Wikipedia for without images in around 31 GB. You can also download a pre-vectorized version. I would also suggest using a SOTA model with deep research, tell it what you're doing, and ask it to compile a report of all the information you'll need to know about clamming and camping in that area, then use RAG on that

[-]

adelaide_flowerpot@reddit

Download a copy of Wikipedia

[-]

AnticitizenPrime@reddit

More parameters = more world knowledge. I agree with others that RAG would be the way to go here. Small models suitable for a phone just don't have the ability to hold much world knowledge.

However, it would be interesting to see small models exclusively trained on a single domain. For example a 4b-9b sized model trained almost exclusively on outdoorsy/survival stuff like you're asking about. That way you could make the most of the parameters. I think this could have a lot of utility.

I don't expect anyone to make a 'clam expert' model anytime soon, though.

[-]

sommerzen@reddit

On my private german knowledge benchmark these models score a bit better than Gemma 2 2b it:

Gemma 3n e4b it (ai studio API, Local model will be probably worse because it's only available at 4bit): 31.7%
Qwen3 4b (no-thinking chat template, q8_K_XL): 28.8%
Phi 4 mini (instruct-5_K_M): 27.9%
Gemma 3 4b it (qat, q4_0): 26.9%
Gemma 2 2b it (q8_0): 25.0%
Llama 3.2 3b it (q8_0): 21.2%
Granite 3.3 2b it (q8_0): 17.3%

Keep in mind that you can't really trust the informations you get from these models when used without rag or something.

[-]

sommerzen@reddit

On my private german "knowledge" benchmark these models score a bit better - Gemma 3 4b (qat, q4_0) -

[-]

Monkey_1505@reddit

I think if you had a really flagship Samsung or Apple phone you might be able to pull this off if you could also have the LLM access a RAG with survival information.

BUT not sure how easy the latter is to set up (probably not), nor am I sure about which models would prompt process fast enough for this even on a flagship phone with fast ram. (Might be 4b size, might be 7b, IDK).

Probably a doable thing, but not with a model in itself, as smaller models are not as good with knowledge, and only with a phone well suited to the tast.

[-]

GortKlaatu_@reddit

Personally I have like three different applications on iOS, since local LLM application developers are slow with updates, and then in each of those I have a variety of the latest popular models.

When I have no connectivity, I may as the same prompt to several of them.

[-]

Ok-Recognition-3177@reddit

I understand the desire, but you would be far better served by visiting Kiwix.com and dowloading a comrpessed offline searchable survival wiki

[-]

Chromix_@reddit

Why not combine both? Someone just released a tool to make offline kiwix searchable for local LLMs. Not sure how much effort it requires to get it running on a phone though.

[-]

Ok-Recognition-3177@reddit

I just saw that, that's such a good idea

[-]

Duxon@reddit

I was playing with Gemma 3n 4b a bit in exactly your scenario. It can identify that a flower is a flower, and a lizard is a lizard, but will confidently give you a wrong and more precise species description.

For language tasks and some general purpose knowledge, I found it really useful for a hike, though. I translated Spanish information boards with it for instance. And those translations were largely spot on.

[-]

westsunset@reddit

Similar for me. It knows a snake is a snake or a lizard a lizard, but it's taking swings at species. if you feed it additional context it might do better. For me though it narrowed the choices to gopher snake or garter snake which don't look alike. Fwiw it was a gopher snake

[-]

FullOf_Bad_Ideas@reddit

Try Qwen 2.5/2 7B VL, it runs in MNN-Chat app. I asked it for recipe of things it can make out of my fridge and it failed massively, so temper your expectations, but it's not useless. Depending on your phone model the best choice will differ - you didn't share how much RAM your phone has.

[-]