Skip to content

Qwen

Fine-tune Qwen3.5 and Qwen3.6 MoE on your data

Qwen's mixture-of-experts models that read images and text, with linear attention in most of their layers. Supported by the training and serving software Tensorant runs. It hasn't run on Tensorant yet, so start with a trial.

Status
Supported
Input
Text and images
Kind
Mixture of experts
Trains with
LoRA
Suggested sizes
LoRA from1 GPU, about 91 GB

Sizes

Sizes and GPUs

Each size with the GPUs one machine needs to train and to serve it. The figures are the estimates Tensorant checks a model against; a trial measures the real use. See GPU memory.

Qwen3.5 and Qwen3.6 MoE models and the GPUs each needs
ModelParametersTrain with LoRATrain with QLoRAServeStatus
Qwen/Qwen3.6-35B-A3B35B, 3B active1 GPUabout 91 GB-1, 2, 4 or 8 GPUsabout 89 GB in allSupported
Qwen/Qwen3.5-35B-A3B35B, 3B active1 GPUabout 91 GB-1, 2, 4 or 8 GPUsabout 89 GB in allSupported
Qwen/Qwen3.5-122B-A10B122B, 10B active--2, 4 or 8 GPUsabout 305 GB in allSupportedServing onlyNeeds its experts split across 8 GPUs, and the trainer has no tested LoRA setup for that on one machine yet.
Qwen/Qwen3.8-2.4T-A95B2.4T, 95B active---Not availableNeeds more than one machine of 8 GPUs.
  • Training needs the memory shown on each GPU: more GPUs on a machine train faster, but each holds the whole model.
  • Serve: the numbers of GPUs of one machine the base model can be split across, and the memory it needs in all.

Training

How it trains

Train Qwen3.5 and Qwen3.6 MoE with LoRA, by supervised fine-tuning (SFT) or DPO on preference pairs.

  • QLoRA isn't offered for this family: choose LoRA.
  • SFT trains with LoRA, on a version with images or of text alone. On text, the vision tower stays frozen and the served model still reads images.
  • DPO trains with LoRA, on one GPU.
  • RL trains text models only.
  • Most of their layers use linear attention. Adapters train those layers and the full-attention layers, not the experts.

Every training run, trials included, pays a training-token fee for the tokens it trains on, by the model's size. Images are on every organization.

Serving

How it serves

Publish a trained model under a name and call it through the OpenAI-compatible API, with the OpenAI SDK or any HTTP client. The name is what your application sends as model.

  • A LoRA adapter you train serves on top of its base model, on the GPUs in the Serve column.
  • The base model serves on its own too, in a test session or for a baseline. A deployment splits it across the GPUs of one machine: the Serve column lists the numbers each size allows.
  • A served vision model reads images and text in the same chat request.
  • Qwen/Qwen3.5-122B-A10B serves as a base model but doesn't train: the table says why.

Steps

Fine-tune Qwen3.5 and Qwen3.6 MoE on Tensorant

  1. Create a project

    Create a project with Qwen/Qwen3.6-35B-A3B as its base model, or pick another size from the list. The model check reads the model before anything starts.

  2. Import your data

    Upload files, paste text or import a Hugging Face dataset as sources.

  3. Review and freeze a version

    Review the examples and approve the ones to train on, then freeze them into a version with training, validation and test splits.

  4. Score the base model

    Evaluate the base model on the version's test split. This baseline is what a training run is compared with, and a full run needs one.

  5. Run a trial, then train

    A trial trains a small sample to check the settings, the GPU memory and the speed. Then start the full run. LoRA on Qwen/Qwen3.6-35B-A3B needs 1 GPU, about 91 GB.

  6. Compare and publish

    Evaluate the trained model on the same test split and compare it with the baseline. Then publish it and call it through the OpenAI-compatible API.

Related

See every model

Start with one dataset.

Create an account and import your data, or book a demo and we will show you the product.