350m vs 685b: liquid ai's tiny phone model takes on deepseek v3

small local mobile model by @liquidai vs deepseek v3 on everyday tasks

we ran a test: lfm 2.5-350m (local run on iphone 14) vs deepseek v3 (inside the deepseek mobile app). the size difference is almost ridiculous

lfm 2.5-350m:
• 361.7 mb (q8_0 quantization)
• 350 million parameters

deepseek v3:
• around 689 gb in model size
• 685b parameters (37b active)

deepseek v3 is roughly 1957x larger by parameter count than lfm 2.5-350m

and yet… for everyday ai tasks, the gap was surprisingly small!

- translation quality: lfm 2.5-350m: 9.5/10, deepseek v3: 5/10
- invoice extraction: lfm 2.5-350m: 8.5/10, deepseek v3: 7/10
- blood test interpretation: lfm 2.5-350m: 4/10, deepseek v3: 8/10

these grades were estimated by gpt 5.5 based on the outputs. they are opinions, not official benchmarks

for many practical tasks, the tiny local model was surprisingly competitive, while deepseek still had a clear advantage on complex interpretation

the gap between open models and closed models is shrinking. and for everyday ai tasks, that gap is already small enough that customer-focused ai companies like openai should be paying attention

the future of mass-market ai might not be everyone renting access to giant cloud models. it might be billions of people running ai directly on their phones