MoonshotAI: Kimi K3 Performance
Compare measured TTFT, throughput, uptime, and route health for MoonshotAI: Kimi K3 across TrustedRouter providers using metadata-only production probes.
moonshotai/kimi-k3
Measured performance
Continuously sampled p50/p95 time-to-first-token (TTFT), effective throughput, and success rate for MoonshotAI: Kimi K3. Effective throughput uses provider-reported output tokens over complete request time. Unsupported route and probe-configuration rows are separated from provider downtime, and no prompt or output content is stored.
| Provider | p50 TTFT | p95 TTFT | Effective throughput | Uptime | Config excluded | Availability samples |
|---|---|---|---|---|---|---|
| relace | 2394 ms | 8967 ms | — | 100.00% | — | 11 |
| kimi | 7274 ms | 19710 ms | 55 tok/s n=1 | 100.00% | — | 238 |
| telnyx | 861 ms | 3246 ms | — | 100.00% | — | 2 |
| wafer | 1463 ms | 1463 ms | — | 100.00% | — | 1 |
| together | 1535 ms | 1686 ms | 89 tok/s n=1 | 100.00% | — | 4 |
| sail-research | 1863 ms | 3538 ms | — | 100.00% | — | 5 |
| makora | 1897 ms | 1897 ms | 13 tok/s n=1 | 33.33% | — | 3 |
| venice | 2002 ms | 2002 ms | 95 tok/s n=1 | 100.00% | — | 1 |
| engy | 2122 ms | 3886 ms | 109 tok/s n=1 | 100.00% | — | 3 |
| io-net | 2147 ms | 2147 ms | 192 tok/s n=1 | 100.00% | — | 1 |
| nebius | 2443 ms | 2443 ms | — | 100.00% | — | 1 |
| phala | 2545 ms | 4076 ms | 33 tok/s n=1 | 100.00% | — | 8 |
| arcee | 3334 ms | 4939 ms | 40 tok/s n=1 | 100.00% | — | 3 |
| fireworks | 4072 ms | 48997 ms | 44 tok/s n=1 | 100.00% | — | 7 |
| featherless | 4823 ms | 4823 ms | — | 100.00% | — | 1 |
| cloudflare-workers-ai | 5457 ms | 5457 ms | — | 100.00% | — | 1 |
| siliconflow | 6170 ms | 13337 ms | 45 tok/s n=1 | 100.00% | — | 3 |
| gmi | 10391 ms | 12634 ms | 42 tok/s n=1 | 100.00% | — | 3 |
| baseten | — | — | — | 100.00% | — | 2 |
| deepinfra | — | — | — | 100.00% | — | 1 |
| tinfoil | — | — | — | 66.67% | — | 3 |
| novita | — | — | 23 tok/s n=1 | — | — | 0 |
| nvidia-nim | — | — | — | 0.00% | — | 2 |
Full provider & model leaderboard.
24 routes.
More routes give the auto router more room to fail over around provider 429 and 5xx responses.
Gateway overhead is measured separately.
Public status separates TLS/health overhead from full model latency so slow LLMs do not inflate the router metric.
Metadata rollups.
Status samples store latency, outcome, provider, model, route, cost, and region metadata only.
View public status or inspect provider routes.