The practical self-host sweet spot — MIT, 1M context, fits 2× H200 with room for full context. Its 1M KV cache is only ~10GB thanks to sparse attention.
Test results
⚡Ping
Latency & availability — single-word reply
1363 ms
🧮Reasoning
Basic math reasoning — show work, give answer
993 ms
{ }JSON Output
Structured output compliance — valid JSON with required keys