June 18, 2026

A broad quality-of-life release across the serving flow and the host console.

Serverless

Model registration redesigned — registering a Hugging Face repo now auto-detects the model’s capabilities, and every derived value can still be overridden.
FP8 / FP4 (NVFP4) quantized models are supported with hardware-aware GPU matching — GPUs that can’t run a model’s quantization are excluded from candidates, with the reason shown.
Staged deployment progress — while a deployment starts you see each phase (pulling image → downloading weights with live % → warming up), and failures carry precise diagnostics: VRAM-vs-RAM out-of-memory classification and a step-level timeline.
Price cap changes are seamless — lowering a cap replaces over-cap replicas gradually, one at a time, without the endpoint ever dipping.
/v1/completions — base (non-chat) completions are supported end to end, with a matching Playground mode.
Webhooks — subscription endpoints with delivery inspection and test-send. Read more →
Artifacts — send generated images/videos to your own storage bucket, filter and delete from the gallery, see auto-delete countdowns, and download via presigned links. Read more →

Clearer machine states — the console now distinguishes Offline, Node Not Ready, and Agent Not Ready, with how-long-ago context, and failed port checks show exactly which port with a one-click retry.
Earnings history — a full earnings table with spreadsheet export.
Machine removal policy — removing a machine with active client workloads starts a 48-hour grace period, and long-offline machines are cleaned up automatically with advance warnings.