- LoRA: Trains a small set of adapter weights on top of the frozen base model.
- Full fine-tuning: Updates every weight in the base model.
Choose a method
Use LoRA when:- You’re starting a new fine-tune: LoRA gets you a working model fastest and at the lowest cost.
- You want to ship multiple adapters from the same base: Adapters are small and can be swapped on a single hosted base model.
- You’re tuning style, format, or domain vocabulary: These are the kinds of updates that LoRA handles best.
- The base behavior needs a substantial change: A model that doesn’t know the task you’re training for may need every weight updated, not just an adapter.
- LoRA results plateau below your target: Try increasing
lora_randlora_alphafirst, and if quality still falls short, switch to full fine-tuning.
Set the method on your job
Thelora parameter defaults to True. Pass lora=False (or --no-lora on the CLI) to run a full fine-tune instead. Everything else about the job stays the same.
LoRA settings
For the parameters that tune LoRA itself (lora_r, lora_alpha, lora_dropout, lora_trainable_modules), see the fine-tuning API reference.
Default target modules
When you don’t setlora_trainable_modules, it defaults to all-linear, which applies LoRA to the modules listed for each model in the tables below. To customize, pass a comma-separated list of module names instead.
Each module you list must appear in the model’s allow-list. Whitespace around module names is ignored, but a non-empty value that parses to no modules (for example "," or " , ") is rejected.
Target MoE expert layers
On mixture-of-experts (MoE) models, you can apply LoRA to the expert feed-forward projections instead of the attention projections. Setlora_trainable_modules to the expert modules w_up, w_gate, and w_down. Together uses a compact shared-factor adapter layout across experts, so the adapter stays small even on very large models.
Use expert targeting when your task depends on the model’s domain knowledge (the feed-forward experts) rather than its attention patterns — for example, adapting an MoE base to a new domain or task family.
- Mixtral:
mistralai/Mixtral-8x7B-v0.1,mistralai/Mixtral-8x7B-Instruct-v0.1. - Qwen3 MoE:
Qwen/Qwen3-30B-A3B-Base,Qwen/Qwen3-30B-A3B,Qwen/Qwen3-30B-A3B-Instruct-2507,Qwen/Qwen3-235B-A22B,Qwen/Qwen3-235B-A22B-Instruct-2507,Qwen/Qwen3-Coder-30B-A3B-Instruct. - Qwen3-Next:
Qwen/Qwen3-Next-80B-A3B-Instruct,Qwen/Qwen3-Next-80B-A3B-Thinking. - GLM-4:
zai-org/GLM-4.6,zai-org/GLM-4.7.
What to expect from full fine-tuning
- Supported models: Full fine-tuning is available for a subset of the models that support LoRA. Large mixture-of-experts models, long-context variants, and some vision-language models are LoRA-only. See supported models for the per-model breakdown.
- Smaller batch sizes: Because full fine-tuning updates every weight, it carries a larger memory footprint, so the maximum batch size for a given model is generally smaller than the LoRA equivalent.
- Higher cost: Full fine-tuning trains every parameter rather than the 0.1% to 1% a LoRA job touches, so it consumes more compute and costs more. See pricing for details.
supports_full_training from the model limits endpoint. When it’s False, the model is LoRA-only, and passing lora=False returns a validation error.
Serve your model
How you deploy depends on the method:- LoRA: After the job completes, deploy the merged model on a dedicated endpoint. See deployment.
- Full fine-tuning: The job produces a complete model rather than a compact adapter. Deploy it on a dedicated endpoint, or download the weights for local use. See deployment.