Skip to content

tencent open-sourced a 440mb translator that beats 220gb models

pulse Man squeezes a GPU-style chip dripping yellow liquid into a "1 BIT" microchip — extreme translation model compression.


tencent just open-sourced a translation model that fits in 440mb and beats models 500x its size

how? they compressed a 3.3gb model by storing weights in basically 1 bit instead of 16

most quantization stops at 4-bit. they pushed to 1.25-bit – and somehow didn't lose accuracy

runs fully offline on a normal android phone. no api calls. no data leaving the device. 33 languages. tibetan and mongolian included

if you're building:
• a travel or language learning app — drop in offline translation with no per-call cost
• a chat or email client — translate messages in the background across any app
• anything for low-connectivity markets — it works without internet, forever
• a privacy-first product — user text never touches a server

0:00
/0:10

this is crazy

Scatter plot: Hy-MT 1.25-bit hits 80 FLORES-200 score at 440MB, beating 9.88GB GemmaX2 and matching 65GB+ Qwen3-32B and DeepSeek-V3.2.

Stay in the loop

Get the latest AI news delivered to your inbox weekly

Thanks for subscribing!