Public service data
Real usage, aggregated.
A rolling 30-day view of metered Inferaft traffic: token volume, model mix, cache reuse and response latency.
Last updated
Oct 27, 2026, 7:00 PM UTC
Tokens processed
91.8M
Complete metered requests, 30 days
Requests
2,203
92.6% completion rate
Cache ratio
92.7%
Share of measured input served from cache
P95 first token
3.9s
8 active models
Daily token volume
Input plus output tokens from completed requests, bucketed by UTC day.
Model mix
Which models are actually carrying traffic, with recent momentum and cache reuse.
| Model | 30d tokens | Share | 7d trend | Cache | Requests |
|---|---|---|---|---|---|
| deepseek-v4.1-flash | 67.3M | 73.3% | -100% | 95.8% | 1,343 |
| kimi-k3 | 13.6M | 14.9% | -53% | 87.5% | 219 |
| glm-5.3-flash | 8M | 8.7% | +521% | 86.8% | 251 |
| deepseek-v4-flash | 1.8M | 1.9% | — | 65.8% | 89 |
| glm-5.3 | 560.4K | 0.6% | — | 5.5% | 23 |
| gemma-4-31b | 349.2K | 0.4% | -100% | 47.5% | 60 |
| qwen-3.8-27b | 159.8K | 0.2% | -100% | 4.0% | 5 |
| z-image-turbo | 29,521 | 0.0% | -100% | 0.0% | 213 |
Methodology
Measured at the gateway.
Token totals include input plus output only when complete usage is available. Cache ratio is cached input divided by measured input. Latency percentiles use completed requests.
Daily buckets use UTC. Seven-day trend compares the latest rolling seven days with the seven days before it. The public dataset exposes aggregate request metadata only.
Prompts and responses are not stored in request history, and this page does not publish account, workspace, key or geographic breakdowns.