LoRA vs. full fine-tuning

• 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

• 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 only an adapter.
• LoRA results plateau below your target: Try increasing lora_r and lora_alpha first, and if quality still falls short, switch to full fine-tuning.

​

Set the method on your job

# LoRA (default)
tg fine-tuning create \
  --training-file "<FILE_ID>" \
  --model "Qwen/Qwen3.5-27B" \
  --lora

# Full fine-tuning
tg fine-tuning create \
  --training-file "<FILE_ID>" \
  --model "Qwen/Qwen3.5-27B" \
  --no-lora

from together import Together

client = Together()

# LoRA (default) — lora=True is optional
job = client.fine_tuning.create(
    training_file="<FILE_ID>",
    model="Qwen/Qwen3.5-27B",
    lora=True,
)

# Full fine-tuning
job = client.fine_tuning.create(
    training_file="<FILE_ID>",
    model="Qwen/Qwen3.5-27B",
    lora=False,
)

import Together from "together-ai";

const client = new Together();

// LoRA (default) — lora: true is optional
const loraJob = await client.fineTuning.create({
  training_file: "<FILE_ID>",
  model: "Qwen/Qwen3.5-27B",
  lora: true,
});

// Full fine-tuning
const fullJob = await client.fineTuning.create({
  training_file: "<FILE_ID>",
  model: "Qwen/Qwen3.5-27B",
  lora: false,
});

​

LoRA settings

​

Default target modules

​

Target MoE expert layers

tg fine-tuning create \
  --training-file "<FILE_ID>" \
  --model "moonshotai/Kimi-K2.7-Code" \
  --lora \
  --lora-trainable-modules "w_up,w_gate,w_down"

from together import Together

client = Together()

job = client.fine_tuning.create(
    training_file="<FILE_ID>",
    model="moonshotai/Kimi-K2.7-Code",
    lora=True,
    lora_trainable_modules="w_up,w_gate,w_down",
)

import Together from "together-ai";

const client = new Together();

const job = await client.fineTuning.create({
  training_file: "<FILE_ID>",
  model: "moonshotai/Kimi-K2.7-Code",
  lora: true,
  lora_trainable_modules: "w_up,w_gate,w_down",
});

• Mixtral: mistralai/Mixtral-8x7B-Instruct-v0.1.
• DeepSeek / Kimi: deepseek-ai/DeepSeek-V3.1, moonshotai/Kimi-K2.6, moonshotai/Kimi-K2.7-Code.
• Nemotron (gateless, w_up and w_down only): nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16, nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.

​

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.

from together import Together

client = Together()

limits = client.fine_tuning.model_limits(model_name="<MODEL_ID>")
print(limits.supports_full_training)

​

Continue training from a checkpoint

• Continue it: The new job picks up the same adapter and keeps training it. The result is still a single adapter on the original base model.
• Merge it: Together folds the adapter into the base model, producing a standalone set of weights, and the new job trains on those instead. The original adapter is no longer a separate, swappable artifact.

• The previous job started from a Hugging Face model: The rejection error names the settings that differ.
• The base model supports LoRA training only: These models never produce full weights, so there is nothing to merge the adapter into.

​

Serve your model

• 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.