FLONNO / SERVICE STATUS

API status and history

The model runs on our own GPU system in Switzerland. These are operator-run measurements, with their scope and limits shown below.

Seven-day measurements

10016 checks, 9995 passed. Latest check: 2026-10-21T16:09:07.401Z. Measurements are recent.

Time to first generated event · p50 / p95

982.35 ms / 1445.93 ms

End-to-end output throughput · p50 / p95

62.00 tokens/s / 85.86 tokens/s

Operator-run synthetic inference probes, thinking disabled, output budget 256 tokens. Throughput includes request latency. Not authenticated gateway availability, customer traffic, independent testing or an SLA.

Percentiles describe successful probes only. Failed probes remain in the availability denominator. No observations are counted as successes. This history measures inference probes, not every gateway, authentication, billing or client failure.

Capacity and recovery

The current deployment is one RTX 3090 GPU system with two active generation lanes and a shared 131,072-token KV pool. The runtime allows up to 16 pending requests and a configured pending timeout of 600 seconds. These are configuration limits, not reserved customer capacity or a guaranteed queueing time.

The gateway allows 60 generation requests per minute and 10 concurrent requests per account, shared across keys. Ten admitted requests do not mean ten simultaneous GPU generations. Long contexts, thinking and other workloads affect queueing and speed. The gateway inference timeout is 900 seconds; clients may time out earlier.

There is no automatic fallback to another AI API provider. An unavailable GPU or tunnel can interrupt service. Clients should use bounded retries with backoff and avoid replaying irreversible tool actions. For a larger workload, request a capacity review with expected RPM, concurrency, context and output sizes. Higher limits require a separate agreement; no higher tier or SLA is promised here.

Reproducible agent API evaluation → · Status JSON · Billing and limits