Recommended Models
Which AI models to use for translation, and how to bring your own.
Recommended Models
| Model | Size | VRAM | Best For |
|---|---|---|---|
| Lightweight (Gemma 4 E2B) | ~1.5 GB | ~2 GB | English, lightweight. Text Reading on low-spec PCs. |
| Standard (Gemma 4 E4B) | ~5 GB | ~6 GB | English & European languages. Text Reading + Image Recognition. |
| High Quality (Qwen3-VL 8B) | ~5 GB | ~6 GB | CJK languages (Japanese, Chinese, Korean). Best Image Recognition for CJK. |
| Gemini Flash (API) | N/A | 0 GB | All languages, highest quality. No GPU needed. Free tier available. |
Language-Specific Picks
| Source Language | Text Reading | Image Recognition |
|---|---|---|
| English | Lightweight / Standard | Standard |
| Japanese / Chinese / Korean | Standard | High Quality or Cloud AI |
| German / French / Spanish / etc. | Lightweight / Standard | Standard |
Using Custom Models
Download a GGUF model
Search HuggingFace for "GGUF" models. Q4_K_M quantization offers the best balance of VRAM usage and quality.
Place the model file
Copy the .gguf file to Playto's models directory:
%APPDATA%\playto\models\
For VLM models: add the mmproj file
If the model supports vision, you also need the mmproj (multimodal projector) file. Place it next to the model:
modelname.gguf
modelname-mmproj.gguf
Select in Settings
Go to Settings > Engine and select your model from the dropdown. Any llama.cpp compatible GGUF model with a supported chat_template will work.
VRAM Performance Guide
| Available VRAM | Recommendation |
|---|---|
| 4 GB | Lightweight model. Use Text Reading. Image Recognition may be tight alongside a game. |
| 6-8 GB | Standard or High Quality. Text Reading + Image Recognition both work well. |
| 12 GB+ | Any model. Game and AI run comfortably side by side. |
| No GPU | Cloud AI (Gemini Flash free tier). Zero local VRAM, fast cloud inference. |
Remember: your game also uses VRAM. The numbers above are VRAM available in addition to what your game needs. Reduce GPU Layers in Settings if you run out of VRAM.