Choose a local training model
The adapter runtime supports three checkpoints. The default is Qwen/Qwen3.5-2B, pinned to Hugging Face snapshot 15852e8c16360a2fea060d615a32b45270f8a8fc.
Check your hardware first
Run tt hardware on the machine that will train the model. It checks the runtime, GPU memory, and model cache space, then reports whether each model fits. Nemotron and Muse Glimmer need substantially more memory than Qwen 2B; DGX Spark is a supported host for both.
Model comparison
| Model | Training | Default sequence length | Estimated training memory |
|---|---|---|---|
Qwen/Qwen3.5-2B | Text SFT with LoRA. This is the default model for tt init. | 2,048 tokens | 16 GiB |
nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 | Text SFT with activation checkpointing and LoRA on shared attention and Mamba projections. | 1,024 tokens | 96 GiB |
meta-models/Muse-Glimmer-30B | Text SFT with activation checkpointing. The vision weights stay frozen and unused; training accepts text examples only. | 2,048 tokens | 96 GiB |
Memory figures are the CLI's planning estimates at the default sequence length and batch size. Actual use depends on the training configuration and other processes on the host. Local adapter training and serving require Linux and an NVIDIA CUDA GPU; evaluation can also run on CPU.
Recommended default
Start with Qwen/Qwen3.5-2B, the smallest supported checkpoint. Compare the base and adapter on a separate evaluation dataset before trying a larger model.
These models are for local training. See Agent setup to configure the model that powers the TT agent.