Create your own mini GPT-4o-mini model via distillation - cheaper and faster, but almost as smart
In 5 minutes, set up fine-tuning of GPT-4o-mini on GPT-4o responses right in the OpenAI dashboard - no code, just the Distill button
- 1.Open the Chat Completions page in the OpenAI dashboard and make sure you have saved requests with the gpt-4o model. If there are few, send 10-50 requests to gpt-4o via Playground or API with different inputs for your task. Tag each request with metadata, for example distillation: my-task, to easily filter them later.
Distillation works like this: GPT-4o (teacher) generates high-quality responses, and GPT-4o-mini (student) learns to replicate them via fine-tuning. Metadata helps keep requests from different models separate. If the service is unavailable in your region, use a VPN or proxy.
OpenAI Dashboard ↗About this tool - 2.On the same Chat Completions page, filter the saved completions: select the gpt-4o model and add a filter by metadata distillation: my-task. Select all found records using the checkboxes.
It is important to choose gpt-4o - you want to distill the responses of the smart model, not the mini itself.
OpenAI Dashboard ↗About this tool - 3.Click the 'Distill' button in the top right corner. In the popup, select gpt-4o-mini as the base model, leave the hyperparameters on auto, and click 'Create'.
OpenAI will automatically create a fine-tuning file from the selected completions and start the training job. The process takes from a few minutes to a couple of hours depending on the data volume.
OpenAI Dashboard ↗About this tool - 4.Go to the Fine-tuning page, find your job, and wait for the 'succeeded' status. Copy the new model's name - it looks like ft:gpt-4o-mini-2024-07-18:personal:XXXXX.
You will need the new model's name for API calls or in Playground - this is your optimized mini model.
OpenAI Fine-tuning ↗About this tool - 5.Test the result in Playground: select your fine-tuned model from the list and send a test request from your task. Compare the response with what a regular gpt-4o-mini would output - the quality should be noticeably closer to gpt-4o.
You can also use the Evaluations tool in the OpenAI dashboard for a formal comparison of models by criteria.
OpenAI Playground ↗About this toolPromptYou are a helpful assistant. [Your task-specific system prompt here] User: [Your test input here]
What this prompt doesEnglish prompt for testing in Playground. Replace [Your task-specific system prompt here] with your task's system prompt, and [Your test input here] with a specific user test request. Before doing this, do not forget to select your fine-tuned model from the model dropdown, not the standard gpt-4o-mini.
- APIOpenAI API - usage feeDistillation requires generating requests via gpt-4o (teacher) - these are paid API calls. Fine-tuning gpt-4o-mini is also billed. The exact cost depends on the data volume; you pay per use, without a separate subscription. All prices are on the pricing.openai.com page.
Why today
AI services change fast - interfaces and free limits may differ from what's described.