Open-source AI large models such as DeepSeek and Qwen perform excellently. With the help of tools like Ollama and LM Studio, we can easily build large model services locally and integrate them into various AI applications, such as video translation software.

However, limited by the VRAM of personal computers, locally deployed large models are usually smaller, such as 1.5B, 7B, 14B, or 32B.

The DeepSeek official online AI service uses the r1 model, which has as many as 671 B parameters. This huge difference means that the intelligence of local models is relatively limited, and they cannot be used as casually as online models. Otherwise, you may encounter various strange problems, such as prompts appearing in the translation results, mixing of original and translated text, or even garbled characters.

The root cause is that small models lack intelligence and have weak understanding and execution capabilities for complex prompts.

Therefore, when using local large models for video translation, you need to pay attention to the following points to obtain better translation results:

I. Correctly Configure the API Settings of the Video Translation Software

Fill in the API address of the locally deployed model in the API interface address under Translation Settings --> Compatible AI and Local Large Models of the video translation software. Usually, the API interface address should end with /v1.

• If your API interface sets an API Key, please fill it in the SK text box. If it is not set, just fill in any value, such as 1234, but do not leave it blank.
• Fill in the model name in the Fill in all available models text box. Note: Some model names may have size information, such as deepseek-r1:8b, and the trailing :8b also needs to be filled in.

II. Prioritize Larger and Newer Models

1. It is recommended to choose a model with at least 7B parameters. If conditions permit, try to choose a model larger than 14B. Of course, the larger the model, the better the effect, provided your computer performance allows it.
2. If using the Tongyi Qianwen series models, prioritize the qwen2.5 series instead of the 1.5 or 2.0 series.

III. Uncheck the "Send Complete Subtitles" Option in the Video Translation Software

Unless the size of the model you deploy is greater than or equal to 70B, checking "Send Complete Subtitles" may cause errors in the subtitle translation results.

IV. Reasonably Set the Number of Subtitle Lines Parameter

Set both the Traditional Translation Subtitle Lines and AI Translation Subtitle Lines in the video translation software to smaller values, such as 1, 5, or 10. This can avoid the problem of too many blank lines and improve the reliability of the translation.

The smaller the value, the lower the possibility of translation errors, but the translation quality will also decrease; the larger the value, although the translation quality is better when there are no errors, it is also more prone to errors.

V. Simplify Prompts

When the model is small, it may not be able to understand or have poor instruction following ability. At this time, you can simplify the prompt to make it simple and clear.

For example, the prompt in the default Software Directory/videotrans/localllm.txt file may be more complex. When you find that the translation results are not satisfactory, you can try to simplify it.

Simplified example one:

# Role
You are a translation assistant who can translate the text within the <INPUT> tag into {lang}.

## Requirements

- The number of lines in the translation must be equal to the number of lines in the original text
- Translate according to the literal meaning, do not interpret the original text.
- Only return the translated text, do not return the original text.
- If you cannot translate, please return a blank line, do not apologize, and do not explain the reason.

## Output format:
Directly output the translated text, do not output any other prompts, such as explanations, guide characters, etc.

<INPUT></INPUT>

Translation result:

Simplified example two:

You are a translation assistant, translate the following text into {lang}, keep the number of lines unchanged, only return the translated text, and return a blank line if you cannot translate.

Text to be translated:
<INPUT></INPUT>

Translation result:

Simplified example three:

Translate the following text into {lang}, keeping the number of lines consistent. If you cannot translate, leave it blank.

<INPUT></INPUT>

Translation result:

You can further simplify and optimize the prompt according to the actual situation.

Through the above optimizations, even smaller local large models can play a greater role in video translation, reduce errors, improve translation quality, and bring you a better local AI experience.