Choosing a model in Playto — a light one to play, a smart one to learn

Playto lets you choose from several models. There’s no need to overthink it up front — Playto picks a sensible default for your setup, so it works without touching anything. This guide is for the next step, when you want to tune things to your own taste.

There’s just one idea to hold onto: don’t pick a model by a leaderboard of raw power. Pick it by the job you want it to do.

Playto has two jobs

There are two situations, and they pull in different directions.

Real-time translation while you play — translating the text on screen on the spot. What matters here is being light and fast. It runs alongside the game, so a heavy model steals performance from the game itself.
Deeper learning — word meanings, grammar notes, example sentences. What matters here is being capable. It’s worth waiting a moment for an explanation that actually lands.

You can do both with one model, or use a different model for each. Knowing you can split them makes the choice easier.

A balanced starting set

Right now, these three work well in Playto. Start here and you won’t go far wrong.

Hy-MT2 1.8B — light, fast, translation-focused. Good for real-time translation while you play.
Qwen3-VL-4B — a general model. It translates and also returns meanings and example sentences, and it sits well with CJK (Japanese, Chinese, Korean).
Gemma 4 E2B — also a general model, comfortable with Latin-script languages.

What the numbers mean

The number in a model’s name is its size — its parameter count. “B” stands for billion. A bigger number tends to mean a smarter model, but also a heavier, slower one that uses more memory (VRAM); a smaller one runs lighter. So 1.8B is lighter than 4B, at the cost of being a little weaker on more involved understanding.

The “E” in Gemma’s “E2B” stands for effective. It’s a newer memory-efficient design, so in practice it runs with roughly the lightness of a 2B-class model.

A light model for in-play translation

For real-time translation, a light, fast, translation-focused model is a good fit. In our own use, Hy-MT2 1.8B strikes a nice balance of speed and lightness and works well as your everyday in-play model.

Two things to keep in mind:

Depending on the language pair, the translation can be less stable. When that happens, switch to one of the general models below.
A translation-focused model only translates. It can’t produce detailed word meanings or example sentences. If you often pull up word details mid-play, it’s smoother to choose a general model from the start.

A general model for meanings, grammar, and examples

When you want meanings, grammar, and example sentences, choose a general model (such as Qwen3-VL-4B or Gemma 4 E2B). These translate as well, and also return definitions and explanations. If your way of playing involves opening word details often, make one of these your main model.

There are loose tendencies by language. Qwen-family models tend to sit well with CJK (Japanese, Chinese, Korean), and Gemma-family models with Latin-script languages — that’s a good starting point. But it’s only a guideline; the surest way is to try both and keep whichever feels right for the language you’re playing in.

Play with a light model — saved words fill in later

Here’s a mechanism worth knowing, because it makes the choice easier. Even if you play with a light, translation-focused model (such as Hy-MT2 1.8B), you don’t have to generate the meanings and grammar of the words you save on the spot.

When a session ends, Playto automatically switches to a general LLM and generates the meanings and grammar for the words you saved that session. It happens after you close the game — at a point where the work no longer competes with the game’s performance. The “Word Book backfill” setting handles this, and even if you’re using a translation-focused model like Hy-MT2 or NMT, it switches to a general LLM just for the backfill. The default “Auto” is fine.

So you can keep play light and enjoyable and still have the detailed review material come together afterward. (If you instead want meanings on the spot during play, that’s the separate case that needs a general model running.)

NMT — a backstop for low-spec machines

NMT is the lightweight option for when your machine simply can’t run a local large model. If yours can, an LLM translates better, so reach for an LLM first. Think of NMT as the last resort that’s there when you need it.

In short: pick by the situation

If in doubt, the default is fine. When you want to tune it, this is all you need:

For in-play translation, a light model like Hy-MT2 1.8B. It won’t get in the game’s way.
For meanings and examples, a general model — Qwen3-VL-4B for CJK, Gemma 4 E2B for Latin-script — as a starting point.
Meanings and grammar for saved words are generated automatically after the session, so you can keep play light.
If a language pair’s translation feels unstable, try a different model.

Try a little, and keep what feels right.

Related: if a translation just doesn’t sit right, we cover what to do in When a translation feels off.