Tencent's Hy-MT2 translator lands on OpenRouter at $0.07/M input
The open-weights family is now a cheap hosted API — 33 languages with glossary and style controls, or run the 1.8B offline in 440MB. Plus Gemma tops 1B.
Copy markdown
$0.07 in, $0.30 out — a real LLM translator, priced like a commodity
Hy-MT2-30B-A3B (30B total, 3B active MoE) is now live on OpenRouter via Tencent Cloud at $0.074 input / $0.295 output per million tokens, 8K context. It covers 33 languages plus five Chinese dialect and minority-language pairs, with structured, delimiter-based, contextual, glossary-based, and style-guided modes — so you enforce a term list or house style in the request instead of post-editing.
Or skip the API: the 1.8B runs offline in 440MB
The whole family ships open weights (Apache-2.0) on Hugging Face, ModelScope, and GitHub, and AngelSlim's 1.25-bit quant squeezes the 1.8B down to ~440MB — small enough to run offline on Apple, Qualcomm, and MediaTek mobile chips at roughly 1.5x the prior speed. On-device, no-network translation is now a realistic weekend build.
Quality: reportedly 88–98% of Gemini 3.1 Pro on FLORES-200
Tencent reports the 1.8B / 7B / 30B tiers reach 88.1%, 96.9%, and 98.1% of Gemini 3.1 Pro on the FLORES-200 benchmark, with the 1.8B beating Microsoft and Doubao's translation APIs overall and the 30B matching or exceeding mainstream models on finance, law, and education text. These are vendor-run numbers — sanity-check on your own language pairs before shipping.
Elsewhere: Google's Gemma crosses 1 billion downloads
Google says its open Gemma family has passed 1 billion downloads with more than 100,000 community variants, and launched an 'Awesome Gemma' repository curating deployment recipes and fine-tunes — a useful jumping-off point if you're choosing an on-device base model.