Measured, not claimed.
Every row below comes from the same probe: a 512-token prompt, one request at a time, 20 samples per model, taken from the gateway itself. The number is the median.
| # | Model | First token | Full reply | Tok/s | Score |
|---|---|---|---|---|---|
| 1 | Mythos China | 0.31 s | 2.9 s | 112 | 96 |
| 2 | LARP | 0.44 s | 3.6 s | 98 | 94 |
| 3 | LAZARUS | 0.58 s | 4.8 s | 81 | 91 |
How a score is built.
Latency, 40%
Median time to first token, inverted and normalised across the field.
Throughput, 35%
Completion tokens per second over the full reply, not a burst window.
Stability, 25%
Share of samples that finished without a retry or a truncated stream.
# reproduce a row yourself
for i in $(seq 20); do
curl -s -o /dev/null -w "%{time_starttransfer} %{time_total}\n" \
https://lazarus-th.duckdns.org/v1/chat/completions \
-H "Authorization: Bearer sk-..." \
-H "Content-Type: application/json" \
-d '{"model":"Mythos China","messages":[{"role":"user","content":"count to 200"}],"max_tokens":512}'
done