Open-source AI large language models like DeepSeek and Qwen have shown excellent performance. With tools like Ollama and LM Studio, we can easily set up local large model services 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 typically smaller, such as 1.5B, 7B, 14B, or 32B.

The official DeepSeek online AI service uses the r1 model, which has a massive parameter count of 671B. This huge difference means that local models have relatively limited intelligence and cannot be used as casually as online models. Otherwise, you may encounter various strange issues, such as prompt text appearing in the translation result, a mix of source and translated text, or even garbled characters.

The root cause is the insufficient intelligence of smaller models and their weaker ability to understand and execute complex prompts.

Therefore, when using local large language models for video translation, paying attention to the following points is necessary to achieve good translation results:

1. Correctly Configure the Video Translation Software's API Settings

Fill in the API address of your locally deployed model into the API Interface Address field under the video translation software's Translation Settings --> Compatible AI & Local Large Models. Typically, the API interface address should end with /v1.

• If your API interface requires an API Key, please enter it in the SK text box. If not set, you can fill in any value, such as 1234, but do not leave it blank.
• Enter the model name in the Fill in All Available Models text box. Note: Some model names may include size information at the end, such as deepseek-r1:8b. The suffix :8b also needs to be included.

2. Prioritize Models with Larger Parameter Counts and Newer Versions

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

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

Unless the model you have deployed is 70B or larger in size, checking "Send Complete Subtitles" may cause errors in the subtitle translation results.

4. Reasonably Set the Subtitle Line Count Parameters

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 helps avoid issues with excessive blank lines and improves translation reliability.

A smaller value reduces the chance of translation errors but also lowers translation quality. A larger value can yield better translation quality when successful but is more prone to errors.

5. Simplify the Prompt

When the model is small, it may not understand or have poor instruction-following capabilities. In such cases, you can simplify the prompt to make it clear and straightforward.

For example, the default prompt in the software directory/videotrans/localllm.txt file might be complex. If you find the translation results unsatisfactory, try simplifying it.

Simplification Example One:

# Role
You are a translation assistant capable of translating the text within the <INPUT> tags into {lang}.

## Requirements

- The number of lines in the translation must equal the number of lines in the original text.
- Translate literally, do not explain the original text.
- Return only the translation, do not return the original text.
- If translation is not possible, return empty lines. Do not apologize or explain the reason.

## Output Format:
Output the translation directly. Do not output any other prompts, such as explanations, guiding characters, etc.

<INPUT></INPUT>

Translation Result:

Simplification Example Two:

You are a translation assistant. Translate the following text into {lang}, keeping the number of lines unchanged. Return only the translation. If translation is not possible, return empty lines.

Text to translate:
<INPUT></INPUT>

Translation Result:

Simplification Example Three:

Translate the following text into {lang}, keeping the line count consistent. If translation is not possible, leave it blank.

<INPUT></INPUT>

Translation Result:

You can further simplify and optimize the prompt based on your actual situation.

By optimizing with the above points, even smaller local large language models can play a greater role in video translation, reducing errors, improving translation quality, and providing you with a better local AI experience.