Everything you need to serve, refine and retrain open-source models, without stitching together separate vendors for each stage.
Serve text, code and vision models through a single OpenAI-compatible API, backed by production SLAs. Choose the Base flavor for cost-efficient, high-volume traffic, or Fast for interactive, latency-sensitive workloads — switch between them instantly, no redeploy required.
Base and Fast return identical outputs from the same checkpoint. The difference is batching and compute allocation, not model quality.
Start on shared endpoints with generous rate limits, then move to a dedicated, isolated endpoint when you need reserved capacity.
Native tool use and structured output are available on supported models, so agents get predictable, parseable responses.
Requests are served from our Singapore, Jakarta or Bangkok facilities, kept within Southeast Asia end to end.
Turn production traffic into training data. Data Lab lets you explore your inference logs, filter for the interactions worth keeping, and curate a dataset without exporting raw logs to a separate tool.
Search and filter historical requests by model, latency, user feedback or custom tags.
Build a labeled dataset directly from real traffic, then export it or send it straight into Post-training.
Curation runs inside the same zero-retention boundary as inference — nothing leaves your account.
Curate data from any endpoint you run on Hi Liberty, including third-party fine-tunes you've deployed.
Adapt an open model to your data, then deploy the resulting checkpoint directly to a Hi Liberty endpoint — same per-token pricing, same SLA, no separate hosting step.
Upload a dataset or pull one straight from Data Lab to start a fine-tuning run.
When training finishes, deploy the checkpoint to a live endpoint without a separate export or hosting step.
Fine-tuned endpoints are billed the same way as base models — per token, with no separate hosting fee.
Pair your fine-tune with a guardrail model like Llama-Guard-4-8B to keep outputs within policy.