Availability ranking: DeepSeek V4.1 Flash last 24 hours
Request success rate over the last 24 hours; fake 200s, truncated streams and malformed responses all count as failures. The range is a 95% Wilson confidence interval based on sample size, so routes with fewer samples get wider intervals.
| Rank | Range | Model @ relay | Availability | Samples | First token | Session cost | Data security |
|---|---|---|---|---|---|---|---|
| 1 | 1 | 98.2%±2.3 | 166 | 1.31 s | ¥2.59 | 4 | |
| 2 | 2 | 57.5%±13.6 | 47 | 1.87 s | — | 4 | |
| 3 | 3–4 | 100.0%±3.8 | 47 | 2.12 s | — | 0 | |
| 4 | 3–4 | 87.0%±9.8 | 46 | 3.04 s | — | 0 |
FAQ
How is availability calculated?
Request success rate over the last 24 hours; fake 200s, truncated streams and malformed responses all count as failures. The range is a 95% Wilson confidence interval based on sample size, so routes with fewer samples get wider intervals.
What does “Range” mean?
We compute a 95% confidence interval from each route’s sample size, then count which routes are clearly better or worse to get the possible rank range. A route at 100% with only a few requests has a very wide range.
Why does the same relay appear more than once?
Each row is one “model @ relay” route. Different models and groups at the same relay can perform very differently, so they are ranked separately.