In an unheralded corner, Tencent Hunyuan released a very small model — just 440MB, 1.8B parameters.
Small isn't the selling point. The claim is that it outperforms Tower-Plus-72B and Qwen3 35B in translation.
Why It Matters
A 72B parameter model is 40x the size of 1.8B. If 1.8B translation quality matches or exceeds 72B-level models, it says two things:
Translation task efficiency has been severely underestimated. General large models have massive parameter redundancy for translation — parameters used for code generation, logical reasoning, creative writing, none of which relate to translation. A 1.8B model specifically optimized for translation can achieve extreme compression on this vertical task.
Edge-side translation experience could be reshaped. A 440MB model fits easily on mobile devices for local inference, no internet needed. If WeChat and QQ's built-in translation switches to this model, both speed and privacy see qualitative improvements.
Some are already speculating that WeChat's translation is running this model under the hood. From experience, WeChat translation's speed and accuracy have always been solid — if it's backed by a local model, that makes sense.
Limitations
This is translation-only performance. A 1.8B model can't compete with 72B on general capabilities. It's a dedicated translator, not a general-purpose model.
If you need offline translation capability, this model is worth watching. Especially for edge deployment — phones, tablets, embedded devices, 440MB has virtually no deployment barrier.
Related reading:
Main source: X/Twitter community discussion