So I bought a phone with Snapdragon 8 elite (gen 4) and 24GB ram (Honor magic 7 pro). My experience has been mixed but with solid potential. Hexagon (Snapdragon 8 Elite) NPU and OpenclGPU support and updates have been rolling in fast but still the fastest prompt processing and token generation have mostly been CPU (I would bet that soon enough either NPU or GPU will be faster or more realistically both). CPU has the downside of generating more heat than NPU and GPU inference but overall it's still the fastest currently.

Now there are no phones with 32gb ram without a virtual ram extension which doesn't work with LLM's ofc, so the best you will do is 24gb ram. What can you do with 24gb ram and a smartphone processor though? Quite a lot actually, MOE has been getting quite popular and their Q4 quants of these models are great and fit into the 24GB. My personal recommendation is IQ4_XS and MXFP4_MOE since with what I have tested MXFP4_MOE is quite faster but for the size IQ4_XS can't be beaten. Q4_0 is more optimised but quality wise it's worse than both (subjectively from my own experience). Goes without saying but Q4_K_M is also quite reliable from a speed/quality/size standpoint.

The main models I use currently are Qwen3.6/3.5-35b-A3B (I prefer 3.5), Qwen3-30b-a3b-2507 (Good quality Less ram more ability to run other applications without crashing) Gemma-4-a4b-26b, LFM-24b-a2b, GPT-OSS-20B. The one I don't reccomend the most is GPT-OSS it's way way too censored and too easy to spook into a refusal if your query even hints at something it deems unsafe. All of them are MOE models which makes intelligence quite good and speed also really good. You can try your luck with different quants of these models but i settled on MXFP4 for max speed at a good quality at a good quality and IQ4_XS for the best quality/size so I can fit other apps into ram and not just be using LLM's.

LFM is by far the fastest and smallest model and it's incredibly smart for its size and speed. They should really make more MOE A2b models because this works so so well. Other models I listed are slower but noticeably smarter.

You will get token generation anywhere between about 25 tokens per second (LFM) and about 11 tokens per second (Gemma). Prompt processing speed really needs to improve though. (LFM is 60 and Gemma is 40 tokens per second). Different quants will have different speeds so use this as just what you will get an average from Q4 quants. Any update will probably make it faster and other advancements like MTP will also make it faster I would assume.

I have no idea whether I should write a guide or not but to keep it simple, if you want to try your luck with your device use pocketpal and as a general rule of thumb load models that don't exceed 75% of your system ram. Dense models will be alot slower (14b dense models are way slower than 20-30b moe models)

A quick test shows Q4_K_M of both models is
55 PP 24 TG LFM2-24b-a2b
13 PP 4 TG Phi-4-14b

Also more A2b and A1b models up to 30b total parameters please and thank you! AND LFM 2.5 24b a2b WHEN?

If anyone has any questions or anything they want me to test don't hesitate to ask.