Export Models
Stage: Ship → Export · What it does: export your trained model into the format you need to run it anywhere — your own inference stack, an edge device, a partner's cloud, or a local laptop.
Why it matters
- No lock-in. The weights are a company asset; export them and move.
- The right format for the right runtime: GGUF for llama.cpp/CPU, ONNX for portability, MLX for Apple Silicon, TensorRT-LLM for NVIDIA serving, safetensors for HuggingFace tooling.
| Format | Best for |
|---|---|
| safetensors | HuggingFace ecosystem, TorchServe, most serving stacks |
| GGUF | llama.cpp, Ollama, LM Studio, CPU/edge |
| ONNX | Cross-runtime portability, ONNX Runtime |
| MLX | Apple Silicon |
| TensorRT-LLM | High-performance NVIDIA serving |
How to use
- Ship → Export.
- Pick the run and a target format.
- Launch — the platform runs the conversion (verified: it loads the model and runs a forward pass before shipping the tarball).
- Grab the produced
model.tar.gzURL and deploy it anywhere, or deploy it back on the platform.
Deploying a raw fine-tune? You don't need to export manually — deploying it auto-builds the serving export behind the scenes (see Ship & deploy). Export is for when you want the artifact in hand or in another format.
Real-world examples
🧑💼 SaaS — bring your tuned model to your own GPU
A SaaS company tunes a support model on the platform, then exports GGUF and runs it on their own CPU-only servers for cost reasons. Same model, their hardware, no per-call fees.
🛒 Retail — edge devices without a network
A retailer wants product-category suggestions on in-store kiosks with no reliable WAN. They export a small fine-tune to MLX/ONNX, load it on-device, and get on-the-spot suggestions without any round-trip to a server.
🏥 Healthcare — a model that travels with the data
A research group exports a clinical summarizer as safetensors and hands it to a partner institution to run inside their compliant environment. The weights, not an API, cross the boundary.
🏦 Financial services — format-of-record
A bank needs an auditable artifact: they export safetensors and archive it with the training run ID, so the exact artifact that passed review is reproducible and portable.
Gotchas
- Pick the format by the runtime, not the hype. GGUF is CPU/edge-friendly; ONNX is portable; TensorRT-LLM is for NVIDIA GPU serving.
- Exports run a real verification load before completion — a broken artifact fails at export, not after you've built an endpoint around it.
- Export records show real status (completed/failed) with the actual error when something goes wrong.