It's a tiny 6M parameter text to image model, and it can make recognizable images at 64x64.
Check out the model card for samples, a detailed explanation about it's architecture and training, and a quick start!
Even the low preset is better than Qwen 3.7 plus or Qwen3.6-27B reasoning
I was gonna make a post about running a Xiaomi LLM locally on a Xiaomi phone but then it just started giving itself a task to complete
Does anyone know what's going on, or how to fix it?
It's an MLP architecture with around 500K total parameters.
Top1
Training accuracy: 5.11%
Validation accuracy 4.59%
Detailed Validation accuracy numbers:
Top-1 Acc: 4.59%
Top-3 Acc: 9.44%
Top-5 Acc: 12.68%
Top-10 Acc: 18.53%
The model was trained on a downscaled version of the Imagenet-1k dataset (32x32) for 5 epochs.
I used pytorch for the training and pyarrow for the dataset, all within termux.
Before anyone comes at me for using an MLP instead of a CNN or similar it's mainly because on my phone an MLP was just more stable, and trained 10-30x faster/step (could be my fault but I'm not too sure). This model specifically took around 30 minutes to train (6 minute/epoch)
The training was entirely on the CPU which is a Dimensity 9300+ and I used 4 of the Arm Cortex-X4 cores.
I might make an improved version later on as this one isn't very accurate.
The image I uploaded here is a 2000 x 2000 downscaled version, but the rendered one is 20000 x 20000 (400 Megapixels), 51 megabytes.
It took 6 minutes to render, 30 seconds to encode and save to png.
Chip used is the Dimensity 9300+
I ran Bonsai 27B on my smartphone. Obviously I had to use the Q1 quant, And even then the speeds I got were really mediocre however impressive for a phone (in my opinion)
(GPU layers says 99 but in reality it's actually CPU only)
Phone used was the Xiaomi 14T Pro (dimensity 9300+)
What model do you guys reccomend to run locally on my phone? Basically any model 4b or under should be fine. My SOC is the dimensity 9300+ running on CPU, with 12gb RAM
These are the models I currently have downloaded:
LFM2.5-1.2B-Thinking-MNN
LFM2.5-1.2B-Instruct-MNN
gemma-4-E2B-it-MNN
Qwen3.5-9B-MNN
Qwen3.5-4B-MNN
Qwen3.5-2B-MNN
Qwen3.5-0.8B-MNN
(Qwen 3.5 9b is kinda useless for me since my phone struggles/crashes with that but I still have it)
I used this phone all the way until last year, now it's just a secondary, and so I decided to custom rom it
It's a very budget build, it has a vo-502, 5 inch coaxial speaker, a TPA3116D2 sterio amplifier (only one channel in use), 12v/1a power supply (I know it's underpowered but I don't use amps full power), and it's inside a shoe box
Basicly what I'm asking for is tips on how to improve it as the sound quality isn't the best, especially the bass, it sounds really bad
I'm using the 2.6x telephoto on the Xiaomi 14t pro in RAW mode, and then manually cropping in
Once the OnePlus 16 / 16 Pro (whatever it will be called) comes out im thinking of buying it off of tradingshenzen, do you guys think it's worth it to upgrade to that phone from my current phone (Xiaomi 14t pro)
Mainly this is how I use my phone: social media, photography, gaming like Delta force
My main concern currently with my Xiaomi is the battery life has been really bad lately (6-7hrs on social media, even worse if doing something else), and the zoom, lowlight camera performance especially on the ultra wide is really bad, I also would like to have a snapdragon chip for on-device AI
Do you guys think that it's worth upgrading to the OnePlus 15 / 16 (when it comes out) from the Xiaomi 14T pro?
My main issues with my current phone is it's battery life, lack of snapdragon chip for on-device ai, and I also wouldn't mind a somewhat better camera
My 14T pro's battery life has been much worse over the past couple of weeks, it drained nearly 50% in just 3 hours, and I wasn't doing anything heavy, or any gaming
Has anyone else been experiencing this, is there a fix?