Loading...
Ghost publishes on-device AI benchmark methodology, per-device rows, and budget thresholds as part of the public trust story — inspectable performance, not marketing claims.
Benchmarks
Published rows & thresholds — on-device.
On-device inference measured end-to-end including tokenization. TTFT = time to first token. TPS = tokens per second.
On-device inference measured end-to-end including tokenization. TTFT = time to first token. TPS = tokens per second.
Adaptive evaluation
Different teams read the benchmark story differently. Pick a focus to route into the right operational surface.
Pick a focus — we route the story to the right surface.
Evaluate Ghost for on-device latency, throughput, memory budgets, and real-world runtime fit.
Start with Benchmarks to inspect TTFT, TPS, speech, and memory budgets, then continue into Command Intelligence or Enterprise based on operational stakes.
Week-one signals
Make Ghost relevant for continuous, multi-step institutional command work by binding decision support to constitutional legitimacy and proof.
| Device | Reflex | Workhorse | STT | TTS |
|---|---|---|---|---|
iPhone 17 Pro A19 Pro · 8GB | TTFT: 280ms TPS: 42 RAM: 520MB | TTFT: 850ms TPS: 18 RAM: 2100MB | RTF: 0.15 RAM: 180MB | VTFT: 120ms RAM: 350MB |
iPhone 16 Pro A18 Pro · 8GB | TTFT: 320ms TPS: 38 RAM: 520MB | TTFT: 980ms TPS: 15 RAM: 2100MB | RTF: 0.18 RAM: 180MB | VTFT: 150ms RAM: 350MB |
iPhone 15 Pro A17 Pro · 8GB | TTFT: 400ms TPS: 32 RAM: 520MB | TTFT: 1200ms TPS: 12 RAM: 2100MB | RTF: 0.22 RAM: 180MB | VTFT: 200ms RAM: 350MB |
Related destinations
Benchmarks only matter when they stay tied to Ghost's locality model, routing posture, and evidence system.