Tool · Inference calculator
What can your hardware serve?
Pick the hardware, precision and context. See which open models fit, how many tokens per second they serve, and what each million tokens costs at your rate.
Total GPU memory: 1,440 GB
DeepSeek-V4.1 Flash763B · 16B active · ~763 GB12K256$1.06
GLM-5.3753B · 40B active · ~753 GB12K256$1.12
Qwen3.8 2.4T-A95B2.4T · 95B active · ~2446 GBDoes not fit
Kimi K32.8T · 104B active · ~2780 GBDoes not fit
DeepSeek-V4 Pro1.6T · 49B active · ~1599 GBDoes not fit
Qwen3.6 27B27.8B · ~28 GB · 8× on 1 GPU102K2,048$0.127
gpt-oss-120b117B · 5.1B active · ~117 GB · 4× on 2 GPUs51K1,024$0.256
gpt-oss-20b21B · 3.6B active · ~21 GB · 8× on 1 GPU136K2,048$0.096
Planning estimates from public specs, not benchmarks: decode throughput from memory bandwidth and compute, with the KV cache sized for the context you picked. Real serving usually lands lower. Cost per million output tokens uses your GPU-hour rate and utilization.
Want this as a real cluster, priced and financed?Start a build
Coming next