Transfer knowledge from our largest models to create smaller, faster versions optimized for your specific tasks and production requirements.
Model distillation trains a smaller "student" model to mimic the behavior of a larger "teacher" model. The result is a compact model that retains much of the capability at a fraction of the cost and latency.
Use the teacher model to generate high-quality responses for your task-specific dataset.
Fine-tune a smaller base model on the teacher's outputs, learning its reasoning patterns.
Deploy the distilled model and continuously improve with production feedback.
from mythicdot import MythicDot
client = MythicDot()
# Step 1: Generate training data from teacher
training_examples = []
for prompt in my_prompts:
response = client.messages.create(
model="mythic-4-ultra", # Teacher model
max_tokens=1024,
messages=[{"role": "user", "content": prompt}]
)
training_examples.append({
"messages": [
{"role": "user", "content": prompt},
{"role": "assistant", "content": response.content[0].text}
]
})
# Step 2: Create distillation job
job = client.distillation.jobs.create(
teacher_model="mythic-4-ultra",
student_base="mythic-3-mini", # Smaller base
training_data=training_examples,
suffix="my-distilled-model"
)
print(f"Distillation job: {job.id}")
# Step 3: Use distilled model
response = client.messages.create(
model="mythic-3-mini:distilled:my-distilled-model",
max_tokens=256,
messages=[{"role": "user", "content": "..."}]
)
Run AI on devices with limited compute. Distilled models work great on mobile apps and IoT devices.
Handle millions of customer conversations cost-effectively with distilled support models.
Meet strict latency requirements for gaming, trading, or interactive applications.
Process large datasets economically with distilled models at scale.
Create production-ready models optimized for your use case.