A single-GPU B300 fleet running DeepSeek V4.1 Flash in FP4 with tiered weight residency — attention and experts in VRAM, overflow in RAM, engram state on SSD. One OpenAI-compatible API, live per-token pricing, monthly spend you can actually read.
FP4, 4-bit Blackwell Ultra, tiered residency.
USD, billed per token, mirrored in /v1/models.
Cheapest of ES · FR · UK · US, migrated on price.
Everything the provider review asks for is machine-readable and public.
GET /v1/models returns schema 2.4 with per-token string pricing, quantization, datacenters and capacity.
POST /v1/chat/completions accepts streaming and non-streaming, and always returns usage token counts.
Prompts are not logged, not retained and not used for training. Full disclosure on the privacy page.