Note 1: Remember that you have to download each update manually. GameTranslate does not currently have any auto-update support!

Back yet again with another update. There are a few more things I'd like to sort out before we hit that 0.5.0 milestone. Therefore, what should have been 0.5.0 is now 0.4.95. Hopefully, the next release will already be on Friday morning, CET.

So! 0.4.95. What does it bring to you?

First of all, the biggest entry here definitely is the integration of Llama.cpp. It will allow users to simply download a Llama-compatible LLM of their choosing, import it into GameTranslate, select their source & target language, optionally set a custom body and afterwards get straight to using it with the translation tool.

The Llama build that comes bundled with this version is running on Vulkan, so it should work just fine on both AMD and Nvidia GPU's. But mind you, LLM's do still require good hardware (mainly a beefy GPU) to run inference (translate) efficiently. For reference, I'm sitting on a GTX 980 Ti, which is the top-of-the-line GPU from 2015, and it is in my opinion way too slow to be used with any LLM that I deem good enough quality. However, if you have any high-end GPU from 2018 and forwards, I reckon it should run just fast enough that it won't be annoyingly slow.

If you're on Nvidia - great, you'll likely have a much better time using LLM's with GameTranslate. I recommend that you download the CUDA version of Llama.cpp and replace the Vulkan version.

1. Download https://github.com/ggml-org/llama.cpp/releases/download/b6123/llama-b6123-bin-wi...
2. Go to GameTranslate's install directory
3. Remove 'llama_cpp' folder
4. Extract the CUDA .zip into the install directory
5. Rename it into 'llama_cpp'

Do not do this if you have an AMD GPU - it won't work!

Where can I download LLM models?

The main website you'll find LLMs is at https://huggingface.co/, but it can be fairly difficult to understand which model you ought to download for your purposes. Hell, it's pretty hard even to find the download button. I will drop links to a model with different sizes below that will work out of the box after you've imported it to the app.

• Gemma 3:4b IQ_1_M (1.2GB) (Direct Download)
• Gemma 3:4b Q3_K_M (2.1GB) (Direct Download)
• Gemma 3:4b Q8_0 (4.1GB) (Direct Download)

 https://huggingface.co/unsloth/gemma-3-4b-it-GGUF

Note: They are about 1-5GB in size (the smaller models). If you have a really strong GPU you might be able to go for an even bigger model.

Which languages can I use with these LLMs?

I simply can not answer that unfortunately, you'll have to do the research yourself. LLMs' are complex beasts and it is often not clear exactly which languages they've been trained on. I will definitely keep testing new models in the future and hopefully we will sooner rather than later see more translation fine-tuned LLMs, which would result in much faster and better quality translations. And hopefully, one day, if GameTranslate grows large enough, I will be training and fine-tuning these models myself! :)

Changelog

• Added Llama.cpp translation alternative
• Custom API is now changed to be pure JSON writing to solve the one-dimensional issue
• Fixed crash reports no longer being sent
• Added 'Force automatic selection' checkbox to the General tab to prevent users from unknowingly being stuck in manual model selection mode after temporarily switching to Custom API
•  Fixed translation bug caused by improper JSON formatting when using Custom API
• Fixed 'manual' mode text overflow when translating single words or short sentences
• Fixed 'manual' mode window width snapping to the shortest sentence
• Fixed 'manual' mode jitter caused by height reset
• Desktop/Attached mode should now properly shutdown if tool crashes
• Fix screenshot API 'DXGI' out of bounds crash
• Fix screenshot API 'DXGI' misalignment caused by game window being outside the monitor in 'Attached' mode
• Disabled launch buttons until tutorial is finished