Skip to content

    Fine-tune a model native to your product

    Improve your agent's domain behavior, tool-use reliability, and response style with fine-tunes that deploy back to the same inference path. Serve them on Dedicated GPUs, through the same OpenAI-compatible API.

    FlexAI Fine Tuning Console

    From your data to a dedicated endpoint

    Push data, run the job, pick a checkpoint, serve. Deploy from the console or CLI; here it is in four CLI commands.

    1. 01

      Push your data

      Upload a dataset once. It is versioned and mounted read-only into every run.

      flexai dataset push support-tickets --file ./tickets
    2. 02

      Run the job

      Your training code from your repo, on 8 GPUs, starting from a Hugging Face base model.

      flexai training run support-sft \
        --dataset support-tickets --accels 8 \
        --repository-url https://github.com/your-org/recipes \
        -- train.py --model=<base-model> --out_dir=/output-checkpoint
    3. 03

      Pick a checkpoint

      List the checkpoints the run saved and choose the one to ship.

      flexai training checkpoints support-sft
    4. 04

      Serve it

      A dedicated endpoint on the same OpenAI-compatible API, scaling from zero to four replicas.

      flexai inference serve support-assistant \
        --checkpoint <checkpoint-id> --accels 2 \
        --min-replicas 0 --max-replicas 4 \
        -- --model=<base-model>

    The endpoint runs on Dedicated GPUs. Every command is in the CLI reference.

    What a run supports

    • Start from open, gated or your own models, or a checkpoint
    • SFT, LoRA, QLoRA and full fine-tuning
    • GRPO and DAPO recipes for reinforcement learning
    • Your code with Axolotl, LlamaFactory, TRL or DeepSpeed
    • Data by upload, S3, GCS, R2 or Hugging Face, versioned and mounted read-only
    • Secrets injected at runtime
    • 1 GPU to multi-node, 8 GPUs per node, on NVIDIA and AMD
    • Checkpoints and resume, with logs and TensorBoard

    See it in practice: LegML fine-tuned a 32B legal LLM that outperformed a frontier model with 50% fewer parameters. Read all the case studies.

    Run your first fine-tune

    Supported today

    What you can fine-tune on

    The same supported matrix as the rest of the platform: pick the framework your recipe already uses and the accelerator it should land on.

    Accelerators
    NVIDIAAMD
    Architectures
    CUDAROCm
    Precisions
    FP8NVFP4MXFP4BF16FP16FP32
    Inference engines
    vLLMSGLangTritonCustom
    Fine Tuning & Training
    PyTorchAxolotlLlamaFactoryTRLDeepSpeed
    Observability & Tools
    GrafanaPrometheusTensorBoardWeights & Biases

    More managed AI services: inference, training, and the platform overview.