Custom checkpoints on Dedicated Capacity

Note

This page covers Dedicated Capacity. Models listed here are not available with an Inference API key unless they also appear in the Inference API catalog.

You can serve a fine-tune of a supported architecture on dedicated capacity.

How it works

1. You fine-tune the model on GPUs, with your own tools.
2. You hand us the weights in Hugging Face format.
3. We serve the checkpoint on your deployment.

We do not fine-tune models.

Requirements for your checkpoint

Your checkpoint must be a fine-tune of a supported model. You can change the weights, but not the architecture.

• Can change: the weights. The tokenizer and the vocabulary can also change, as long as the vocabulary size stays exactly the same. This helps for multilingual models.
• Must stay the same: the number of attention heads, the RoPE settings, the vocabulary size and the other architecture values in config.json.
• Format: Hugging Face format, with the weights as safetensors files and the same config and tokenizer files as the base model.

For example, DeepSeek-R1-Distill-Llama-70B and Llama-3.3-Swallow-70B-Instruct-v0.4 are fine-tunes of Meta-Llama-3.3-70B-Instruct. They keep its architecture, so they can run on dedicated capacity.

Which models accept custom checkpoints

The Custom checkpoints column on Supported models shows which models accept them.

These models do not accept custom checkpoints:

• Multimodal models, for example Gemma 4 and MiniMax M3.
• Models without a conversion package yet, for example GLM-5.3.
• Models that do not generate text, for example speech and embedding models.

Call your checkpoint

• Your checkpoint gets its own model ID. You call it through the same API as every other model. See Connect to your endpoint.
• It can run in a bundle next to other models, for example next to its base model. See Model bundles.
• If your checkpoint uses a different chat template or tool-call format than its base model, handle the prompt format and the output parsing in your client.

Background

SambaNova describes how Hugging Face checkpoints are converted for its chips: checkpoint conversion tool.

Support for a new architecture is a commercial conversation. Talk to us.