Skip to main content
You can train the following models:
  • OpenPipe/Qwen3-14B-Instruct
  • Qwen/Qwen3-30B-A3B-Instruct-2507
  • meta-llama/Llama-3.1-8B-Instruct
The best model depends on your task. A smaller model might be faster and cheaper, while a larger model might handle more difficult behavior. Model Distillation lets you train several candidates and compare them on the same examples. The UI lists a model only while the training service accepts it, so a model can temporarily disappear from the selector.

​
Experimental models

The following models are available on request as experimental training targets. They might not appear in your model selector by default, and support can change without notice.
  • nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
  • Qwen/Qwen3.5-4B
  • Qwen/Qwen3.5-9B
  • Qwen/Qwen3.6-35B-A3B
  • Qwen/Qwen3.8-27B
  • google/gemma-4-26B-A4B-it
  • google/gemma-4-31B-it
The UI shows the Nemotron model as nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B. In API requests, use the full identifier, including -BF16.
These lists apply to models you can train. You can also connect existing models from OpenAI-compatible providers for serving, relabeling, and comparison.

​
You do not need to choose one model in advance

Train a few promising models, evaluate them against the behavior you want, and deploy the strongest result. This lets you compare results for your task instead of choosing a model based only on benchmarks or size.
When you create a fine-tune, copy a listed value exactly into base_model:
Use the complete request in Fine-tuning Quick Start.
Last modified on September 21, 2026