Manual Installation of AI Model Weights
If the automatic download of AI model weights failsβdue to internet restrictions or other issuesβyou can manually download and install the necessary files. This guide explains where to place the files and how to create the required directories if they do not already exist.
First, check config.json to confirm which models are currently configured before downloading files.
See the Check config.json section below.
AI Models Directory Structureβ
Below is the directory structure of VRCT:
C:\Users\{YourUsername}\AppData\Local\VRCT
(or wherever the folder containing VRCT.exe is located)
βββ VRCT.exe
...
βββ weights
βββ ctranslate2
β βββ {CTRANSLATE2_WEIGHT_TYPE} β folder name must match config.json exactly
β βββ config.json
β βββ model.bin
β βββ ...
βββ whisper
βββ tiny
βββ base (Required by default)
β βββ config.json
β βββ model.bin
β βββ tokenizer.json
β βββ vocabulary.txt
βββ small
βββ medium
βββ large-v1
βββ large-v2
βββ large-v3
The CTranslate2 folder name must exactly match the value of CTRANSLATE2_WEIGHT_TYPE in your config.json.
For example, if config.json contains "CTRANSLATE2_WEIGHT_TYPE": "nllb-200-distilled-600M-ct2-int8", the folder must be named nllb-200-distilled-600M-ct2-int8.
Check config.jsonβ
Before downloading anything, open your config.json and confirm these two values:
"CTRANSLATE2_WEIGHT_TYPE": "nllb-200-distilled-600M-ct2-int8",
"WHISPER_WEIGHT_TYPE": "base",
CTRANSLATE2_WEIGHT_TYPEβ determines the CTranslate2 (translation) model to use and must match the folder name underweights\ctranslate2\.WHISPER_WEIGHT_TYPEβ determines the Whisper (transcription) model to use and must match the folder name underweights\whisper\.
Make sure you download and place the files that match these values.
Model Download URLsβ
VRCT needs a CTranslate2 (translation) model and a Whisper (transcription) model.
CTranslate2 (Translation)β
The default translation model is nllb-200-distilled-600M-ct2-int8.
Place all downloaded files into weights\ctranslate2\{CTRANSLATE2_WEIGHT_TYPE}\.
Clicking the direct download links below will immediately start a download.
NLLB-200 Distilled 600M (default) β place in weights\ctranslate2\nllb-200-distilled-600M-ct2-int8\
- config.json
- model.bin
- sentencepiece.bpe.model
- shared_vocabulary.txt
- special_tokens_map.json
- tokenizer.json
- tokenizer_config.json
Alternatively, browse the JustFrederik/nllb-200-distilled-600M-ct2-int8 model page on Hugging Face.
Whisper (Transcription)β
Download the specific Whisper model that matches your WHISPER_WEIGHT_TYPE setting. The base model is required by default.
Place all downloaded files into weights\whisper\{WHISPER_WEIGHT_TYPE}\.
| Model | Link | Notes |
|---|---|---|
| tiny | Hugging Face | |
| base | Hugging Face | (Required by default) |
| small | Hugging Face | |
| medium | Hugging Face | |
| large-v1 | Hugging Face | |
| large-v2 | Hugging Face | |
| large-v3 | Hugging Face |
For each Whisper model, download these 4 files:
config.jsonmodel.bintokenizer.jsonvocabulary.txt
For Whisper models from the Systran series, only the above 4 files are needed.
preprocessor_config.json and vocabulary.json are not required.
Direct download links for the base model:
Still having issues?β
-
Check the folder name β Make sure the CTranslate2 folder name under
weights\ctranslate2\exactly matchesCTRANSLATE2_WEIGHT_TYPEinconfig.json. A mismatch is the most common cause of failure. -
Check VPN / Proxy / Antivirus β If automatic download is failing, verify that your VPN, proxy, or antivirus software is not blocking connections to
huggingface.co. -
Delete and re-download β Delete the
weights\whisperandweights\ctranslate2folders and let VRCT re-download them on the next launch. -
Attach logs when reporting β If the issue persists, attach your logs and settings data when filing a GitHub issue.