Hy-MT2 1.8B in practice: what the model is, and how it behaves in a live game overlay
Hy-MT2 1.8B is the translation model Playto installs by default. This is the practical write-up: what the model is, the settings we run it with, what we measured on game text, and where it stopped working for us. The wider comparison with five other local models is in Best local LLM for real-time translation, and the plain-language guide to picking a model is Choosing a model.
What Hy-MT2 is
- A family of translation-only models from Tencent, released 21 May 2026 under Apache 2.0.
- Three sizes: 1.8B, 7B, and a 30B mixture-of-experts. Everything below is about the 1.8B; we have not measured the larger two.
- One model covers 33 languages, including Japanese, Chinese, Korean, the major European languages, Vietnamese, Turkish, Arabic and Hindi.
- Translation only. It does not answer questions, define words or write example sentences. That is a feature for a live overlay (nothing to go wrong) and a limit for a learning tool (meanings need a general model).
- Tencent publishes GGUF files, including experimental 2-bit and 1.25-bit builds.
How Playto runs it
- Quantization: Q4_K_M, 1.13 GB on disk. Since the August build the file ships inside the installer, so a fresh install translates with no download.
- Sampling: temperature 0.7, top-p 0.6, top-k 20, repetition penalty 1.05, which are the values on the model card.
- Prompt: the one-line instruction from the model card, one line of game text at a time. No system prompt, no JSON.
- Default for new installs when both languages of the pair are in the 18 we verified (ja, en, zh, zh-Hans, zh-Hant, ko, de, fr, es, pt, it, pl, vi, ru, el, tr, cs, nl). Other pairs default to a general model. A model you chose yourself always wins.
What we measured
Behavior on 72 game lines (9 English and 9 Japanese lines of dialogue, UI and system text, each into 4 languages), RTX 4070 Ti, llama.cpp server build 8724:
| Hy-MT2 1.8B Q4_K_M | |
|---|---|
| Right target language | 69 / 72 |
| Prompt echo or source passthrough | 0 |
| 3-line block returned as 3 lines | 3 / 3 |
| Median per short line | 85 ms |
| Decode speed | 221 tok/s |
| VRAM in use | 1.77 GB |
Quality on a 50-line English–Japanese reference corpus, both directions, chrF against reference translations:
| Hy-MT2 1.8B | Opus-MT (classic NMT) | Gemma 4 E2B | |
|---|---|---|---|
| chrF, mean | 54.1 | 35.5 | 61.6 |
| Japanese to English | 62.8 | 46.9 | 70.8 |
| English to Japanese | 45.4 | 24.1 | 52.4 |
| Short dialogue | 54.8 | 40.7 | 61.1 |
| Skill descriptions | 53.8 | 32.3 | 60.9 |
| Proper nouns | 51.0 | 36.5 | 56.1 |
Reading the outputs: Hy-MT2 is consistent and slightly literal. It rendered “magical damage” where a human writes “magic damage”, turned an undead army into a “corpse army”, and once invented katakana for a term (コールドダウン for cooldown). It never dropped a line, never explained, and never switched language on us. For a model that has to answer in 85 ms while the game is running, that reliability is the point.
How small you can go
We tried to get the 1.1 GB file under 1 GB with the usual quantization recipes, measured on CPU with the same 72 lines:
| Build | Size | Right language | Prompt echo | Median per line (CPU) |
|---|---|---|---|---|
| Q4_K_M | 1081 MB | 69 / 72 | 0 | 0.93 s |
| IQ3_M | ~790 MB | 61 / 72 | 19 | 0.91 s |
| IQ3_XXS | 733 MB | 67 / 72 | 38 | 1.05 s |
| Q3_K_M | ~860 MB | 53 / 72 | 12 | 0.97 s |
| IQ2_S | 657 MB | 31 / 72 | 35 | 0.90 s |
Below 4 bits the model starts repeating the request instead of translating, even with a game-text calibration set, and it is not faster. Use Q4_K_M. Tencent’s own 2-bit and 1.25-bit files use a newer quantization format that, when we checked in August, needed a llama.cpp branch not yet available for mainstream x86 builds.
Two things that did work:
- CPU fallback. Q4_K_M with no GPU offload gives identical output at about one line per second on a desktop CPU. Too slow to feel live, fine when the GPU is unavailable. Playto uses this as the fallback when the GPU engine cannot start.
- Language pairs. The coverage claim is 33 languages; in our 20-sentence checks Greek, Turkish, Czech and Dutch were fine, while Romanian, Danish, Norwegian, Finnish, Hungarian and Swedish broke often enough that Playto routes those pairs to a general model instead. Check your own pair before relying on it.
Hy-MT2 or Gemma 4 for translation?
Both are worth having, for different reasons.
| Hy-MT2 1.8B | Gemma 4 E2B | |
|---|---|---|
| chrF on our corpus | 54.1 | 61.6 |
| Download | 1.1 GB (bundled) | 3.5 GB |
| VRAM in use | 1.8 GB | 2.3 GB |
| Median per short line | 85 ms | 96 ms |
| Word meanings, examples | No | Yes |
| Behavior on 72 lines | Clean | Clean |
Gemma 4 E2B is the better translator and the only one of the two that can explain what you saved. Hy-MT2 is a third of the download, starts in seconds on machines without a dedicated GPU, and leaves half a gigabyte more for the game. That is why Playto ships Hy-MT2 as the default and offers Gemma 4 E2B and Qwen3-VL 4B as one-click upgrades, and why saved words get their meanings generated after the session rather than during play.
Summary
- Hy-MT2 1.8B is a reliable, fast, small translation-only model. On game text it is far ahead of classic NMT and a step behind a 2B-class general model.
- Run Q4_K_M with the model card’s sampling values. Do not go below 4 bits.
- Verify your language pair; six of the claimed pairs did not hold up for us.
- If you have 2.3 GB of VRAM to spare and want meanings, pick Gemma 4 E2B. If you want the smallest thing that works everywhere, this is it.