Frontier open models served on B300 clusters by the engineers behind Mooncake, KTransformers and SGLang — fast, OpenAI-compatible, and priced by the token.
base_url = "https://api.cantusai.io/v1"Each model is deployed, tuned and benchmarked by our inference team before it goes live. Swap models by changing a single string.
| Model | Input / 1M | Output / 1M |
|---|
A KVCache-centric serving stack: prefill and decode run on separate pools, reusable context lives in a shared cache, and the scheduler routes every request to where its prefix already is.
KVCache-centric architecture that splits prefill and decode, and pools cache across GPU, DRAM and SSD to raise throughput under SLOs.
CPU/GPU co-execution for giant mixture-of-experts models — placing hot and cold experts where they cost the least.
High-performance serving runtime with RadixAttention prefix reuse, zero-overhead scheduling and structured outputs.
Multi-region routing, rate limiting and metering on KYC-verified GPU capacity — every token counted, every request traced.
Call any model in the catalog. No GPUs to manage, no minimums.
Your model on reserved B300 capacity, isolated and tuned to your traffic shape.
Committed token volume at a locked price for platforms and enterprises building on open models.
In polyphony, one line moves steadily forward while every other voice is composed around it. We build that line for AI: a steady stream of tokens that everything else can be built on.
Latin cantus, from canere, “to sing” — a song, a melodic line, a voice. The cantus firmus is the line that carries the piece.
Tokens leave the factory one by one along a single line — each with direction, each counted. We sell the flow, not idle hardware.
KYC-verified GPU capacity, a world-class inference team, and global distribution — integrated into one token factory.
API access, dedicated endpoints and reserved throughput.