Recommended
The sensible default
The best balance of capability, privacy, and money for the constraints you chose.
- Hardware
- ~$5,499
- Cloud / month
- $14–$23
- Runs locally
- 80%
- Breaks even
- Tokens winThe hardware never recoups its cost at this volume. Tokens stay cheaper.
MachineM5 Ultra Mac Studio96 GB unified · 1.2 TB/s
A quiet, turnkey complete system at list price for capacity-first MLX or llama.cpp serving.
Apple Mac Studio newsroom ↗Model routes
DeepSeek V4 Flash 0731 (experimental Q2)local
DeepSeek V4 Flash 0731 (experimental Q2)localcoding + research + document work
~144 tok/s · bandwidth estimate
source ↗
Amazon Nova Lite
Claude Haiku 5.5
GPT-6 LunaWhy it fits
- Separates capacity-heavy and latency-sensitive local jobs across complementary hardware.
- Routine work stays on the economical route while difficult jobs have a clear fallback.
- Ranked on a measured intelligence index of 43.
Reality check
- Large DeepSeek quantizations on Spark are experimental and need hands-on serving work.
- Treat hardware prices as planning allowances and check a live listing first.
- API estimate assumes 70% input / 30% output and excludes caching discounts.
- Sustained heavy volume can outrun personal plan limits.
- Token speeds come from memory-bandwidth math at batch size 1. Measure before you rely on them.
Gemma 4 31B
Qwen3.8 27B