Tencent HY-MT2 1.8B Gets a Hosted API: Tencent's 440MB Translator Goes Cloud

Tencent HY-MT2 1.8B, the compact member of the Hy-MT2 translation family Tencent Hunyuan open-sourced on May 21, 2026, is now callable over a hosted commercial API rather than only self-hosted — the model's API listing went live on August 20, 2026, at roughly $0.044 per million input tokens and $0.177 per million output tokens. That listing date matters more than it usually would, because the 1.8B was never just another open checkpoint: it is the model Tencent built to prove that a 440MB quantized translation model can run entirely on a phone, and in the three months since the open-source release it has also been shipping inside Tencent's own HyTranslate mini-program and the iOS and Android apps that followed. This article is about what the hosted-API launch changes, what the model actually is, and which numbers you should trust.

What actually changed on August 20

Until now, using Tencent HY-MT2 1.8B meant one of two things: loading the Apache-2.0 weights into transformers, vLLM, SGLang, or llama.cpp yourself, or using Tencent's consumer translation products. The hosted API is a third path — you send text over HTTP, and Tencent Cloud serves the 1.8B endpoint. The model keeps its 8,192-token context and 4,096-token maximum output, handles text only, and is priced at about $0.044 per million input tokens and $0.177 per million output tokens. For teams whose translation volume is spiky, that removes the "must operate a GPU" floor from trying the model.

The other half of the story is the apps. The HyTranslate mini-program launched alongside the open-source release in May with voice input, style presets, and offline mode. The iOS and Android apps — which download the on-device model for network-free translation — have since shipped. The 1.8B is the workhorse of that offline story: with AngelSlim's 1.25-bit extreme quantization it shrinks to about 440MB of storage and runs locally on Apple, Qualcomm, and MediaTek phone silicon, with Tencent reporting 1.5x faster inference than the previous generation's 4-bit build on an Apple A15.

What the 1.8B is

HY-MT2-1.8B is a dense, decoder-only causal LM built on the hunyuan_v1_dense architecture — about 2 billion parameters in practice despite the 1.8B name — with a chat template and no default system prompt. It translates among 33 languages plus five minority-language and dialect pairs, including Chinese, English, Japanese, Korean, French, Spanish, Russian, Arabic, Thai, Vietnamese, Indonesian, Hindi, Bengali, Tamil, and the Tibetan, Uyghur, Kazakh, Mongolian, and Cantonese dialect pairs. Unlike the classic encoder-decoder translation models, it is instruction-tuned: you prompt it with the target language and optional instructions, and it can preserve JSON and HTML structure, honor a glossary, match a register, and use surrounding context to disambiguate. That capability class is what the sibling 7B and 30B-A3B models share, and what Tencent's IFMTBench benchmark was released alongside to grade.

Where the benchmarks stand

The quality claims are worth reading carefully because almost all of them are Tencent's own, unreproduced by an independent lab as of this writing. Tencent reports HY-MT2-1.8B at 89.9% of Ge​mini 3.1 Pro's average FLORES-200 score (the 7B at 97.9%, the 30B-A3B at 98.6%), at 96.7% of Ge​mini on its real-world GEMBA-style test set, and at 96.2% on DomainMTBench across eight professional domains. Some press coverage rounded the FLORES-200 figure down to 88.1%; the vendor README lists 89.9%. Tencent also claims the 1.8B overall surpasses mainstream commercial translation APIs from providers such as Microsoft and Doubao. Independent reproductions do not exist yet, so treat all of that as vendor-reported directional evidence rather than audited fact. What is not vendor-dependent: the 33+5 language set is fixed, the Apache-2.0 license is real, and a 440MB quantized checkpoint that runs on a phone is independently observable.

How you'd use it, and what it costs

• Self-hosted — Apache-2.0 weights from Hugging Face or ModelScope; run with transformers (>=5.6.0), vLLM, SGLang, or llama.cpp GGUF builds. Cost is your GPU bill; the 1.25-bit build runs on CPU-class devices.

• Hosted API — Tencent Cloud serves the 1.8B at about $0.044 per million input tokens and $0.177 per million output tokens as of this writing; the vendor's own API and several third-party platforms carry it.

• Consumer — HyTranslate mini-program and the iOS/Android apps, including offline translation via the downloaded on-device model.

Recommended sampling for the 1.8B: temperature 0.7, top-p 0.6, top-k 20, repetition penalty 1.05, max tokens 4096.

If you are building a translation pipeline rather than shipping a consumer app, the interesting thing about a model like this is that its price is a moving target — Tencent already cut its Hy-MT2-Pro API rate in June. That is exactly the scenario a 0% markup router exists for: because the pass-through is at provider list price, a vendor price cut shows up on the same-day bill rather than after your gateway reprices. The model itself is not on OrcaRouter's roster at the time of writing, so you would self-host the weights or call Tencent Cloud's API directly; the failover and single-key argument applies to the broader translation stack you'd build around it, where mixing a small on-device model for cheap bulk traffic with a larger model for hard pairs is the kind of composition routing was built to express.

The bottom line

Tencent HY-MT2 1.8B was already interesting in May as a phone-sized Apache-2.0 translator; the August 20 hosted-API listing makes it a candidate for teams that want to evaluate it without standing up infrastructure. The quality numbers are vendor-reported, the language set is deliberately mainstream rather than maximal, and the on-device footprint is the genuinely distinctive thing about it. If your pairs are inside its 33, it is now cheaper than ever to find out whether the claims hold for your content.