- Avg Tokens / Second
- 40.20
- Decode Tokens / Second
- 42.40
- Avg Time to First Token (ms)
- 550.00
- Runs Analysed
- 2
- Last Updated
- Aug 16, 2026, 04:34 PM
Throughput distribution
Throughput over time
Generated throughput, including reasoning tokens — so a thinking model's line does not collapse; the table ranks on visible tokens · one cell per model, showing its best-covered provider · models with at least 4 samples in the window first, fastest of those at the top · shared vertical scale, 0 to 52 tok/s (99th percentile) · dashed rule is the model's own mean · 1 cell carry fewer than 4 samples
Every lane measuring this model
2 measured · live first · a lane is a provider, a catalogue id and a transport| Provider | Lane | Avg Toks/Sec | Min | Max | Avg TTF (ms) | n |
|---|---|---|---|---|---|---|
| deepinfra | via OpenRouter | 40.20 | 31.20 | 49.10 | 550.00 | 2 |
| deepinfra | direct | 33.60 | 32.20 | 35.00 | 590.00 | 2 |
Frequently Asked Questions
How fast is Phi 4?
The latest rolling average throughput is 40.20 tokens per second with an average time to first token of 550.00 ms across 2 recent runs. Once tokens are flowing the decoder delivers about 42.40 tokens per second, measured from the gaps between stream chunks with time to first token and reasoning excluded.
How often are these benchmarks updated?
Benchmarks refresh automatically whenever the monitoring cron runs. The most recent run completed on Aug 16, 2026, 04:34 PM.