Throughput
What one Mac can actually serve.
Cluster capacity starts from measured hardware — not slideware. Size a dedicated cluster when your team outgrows a single node.
Throughput
What one Mac can actually serve
Measured on a single 128 GB M5 Max under sustained load, with simulated users based on real Midium usage.
| Model | Coding developers (Standard) | Chat users (Standard) | Background agents |
|---|---|---|---|
| Qwen3.6 35B | 4 (~6 at Relaxed) | 24 (~40 at Relaxed) | 2 at once, 238k tokens/h |
| Gemma 4 26B | ~5 | 16 | 2 at once, 225k tokens/h |
| Ornith 1.5 122B | 2 | 12 at Relaxed | 4 at once, 134k tokens/h |
Standard: coding 25 tok/s and first token within 5 s; chat 20 tok/s and 2 s. Relaxed: coding 15 tok/s and 10 s; chat 12 tok/s and 4 s.
- Generation stays fast (50–220 tok/s per user). Reading long prompts sets the limit.
- Qwen3.6 stays fast on long 128k-token coding sessions.
- M5 Ultra and M3 Ultra: testing now. Estimates unpublished until measured.
Cluster load
Up to Sonnet-class intelligence on your own Macs.
Up to Sonnet Class Intelligence
Monthly tokens
~709M
~648M in · ~61M out
Same tokens on Sonnet 5.5
~$1,907/mo
$2 / $10 per 1M
128 GB Mac Studio
$1,000/mo flat
1.9× cheaper