seconds checkpoint → endpoint
Fine-tuning that ships
SFT → DPO → RL, then deploy the checkpoint in seconds.
Register your dataset, then launch a managed SFT or DPO job metered per 1M training tokens, or run dedicated RL metered per GPU-second. Progress streams live in the console, and finished checkpoints promote straight to inference — serverless or dedicated — behind the same API shape your application already calls.
SFT and DPO rates follow the parameter band of the base model: up to 16B, LoRA SFT is $0.50/1M training tokens and full DPO $2.00/1M; on the 80–300B band, LoRA SFT is $6.00/1M and full DPO $24.00/1M. The full matrix sits on /pricing.
RL is the shape that needs dedicated hardware: it holds weights, optimizer state, and rollout batches at once, so the bill follows GPU-seconds rather than tokens. The catalog's `b200-4x-fp4` preset supplies 4× B200 180GB on NVFP4 at $40/hr and is the entry shape for ≤70B dense tuning; `b300-8x-fp8` supplies 8× B300 288GB on FP8 at $96/hr when maximum quality on frontier dense weights matters more than tokens per dollar.
No catalog model carries the tunable flag yet, so this page links the GPU shapes rather than a model page. Pick the band your base model falls into on /pricing, confirm the preset has the memory for the job, then launch from the console and promote the checkpoint to the inference endpoint when the eval passes.
- ✓LoRA to full RL spectrum
- ✓Multi-LoRA serving
- ✓Dataset registry built in
- ✓Eval before you promote
Keep exploring
Production inference
Serverless tokens or warm dedicated GPUs — one gateway, OpenAI-compatible.
Agents that stay up
Low-latency tool loops, streaming, and spend guardrails for autonomous workloads.
Batch API (beta)
Evals, backfills, and bulk scoring — async, without holding a connection.
Enterprise RAG that cites
Embed, retrieve, rerank, answer — with cached prefixes and audit trails.
Long-context production
Text models behind one key — vision is not served yet.