Where the estimate matched a measurement, and where it has not been measured
3 in-house calibrations are listed on this page. The compared rows are Qwen2.5-3B-Instruct on GeForce RTX 4090 24GB (desktop). Predicted total bytes are compared with measured process GPU bytes, not with board-used bytes. These rows stay separate and are not rolled into one error figure. Planned, with no measured bytes: Qwen2.5-7B-Instruct, Qwen2.5-32B-Instruct, and NVIDIA DGX Spark (GB10, 128 GB unified).
Measured calibrations
| Calibration | Model | Hardware | Engine | Quantization | Context | Captured | Measured process bytes | Predicted low | Predicted high | Measured minus low | Measured minus high | Evidence path |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
rt-qwen2.5-3b-instruct-q4km-4090-ctx512 | Qwen2.5-3B-Instruct | GeForce RTX 4090 24GB (desktop) | llama.cppversion: 0.4.1-dev (build 1, commit 1af554f) | GGUF Q4_K_Mgguf-q4_k_m | 5121 · fp16 | 2,485,125,120 bytes 2.31 GiB Inside the predicted range | 2,258,024,864 bytes 2.10 GiB | 2,703,154,191 bytes 2.52 GiB | +227,100,256 bytes +0.21 GiB | -218,029,071 bytes −0.20 GiB | docs/evidence/calibrations/rt-qwen2.5-3b-instruct-q4km-4090-ctx512.json | |
rt-qwen2.5-3b-instruct-q4km-4090-ctx2048 | Qwen2.5-3B-Instruct | GeForce RTX 4090 24GB (desktop) | llama.cppversion: 0.4.1-dev (build 1, commit 1af554f) | GGUF Q4_K_Mgguf-q4_k_m | 2,0481 · fp16 | 2,543,845,376 bytes 2.37 GiB Inside the predicted range | 2,314,647,968 bytes 2.16 GiB | 2,760,909,757 bytes 2.57 GiB | +229,197,408 bytes +0.21 GiB | -217,064,381 bytes −0.20 GiB | docs/evidence/calibrations/rt-qwen2.5-3b-instruct-q4km-4090-ctx2048.json | |
rt-qwen2.5-3b-instruct-q4km-4090-ctx8192 | Qwen2.5-3B-Instruct | GeForce RTX 4090 24GB (desktop) | llama.cppversion: 0.4.1-dev (build 1, commit 1af554f) | GGUF Q4_K_Mgguf-q4_k_m | 8,1921 · fp16 | 2,776,629,248 bytes 2.59 GiB Inside the predicted range | 2,541,140,384 bytes 2.37 GiB | 3,007,031,516 bytes 2.80 GiB | +235,488,864 bytes +0.22 GiB | -230,402,268 bytes −0.21 GiB | docs/evidence/calibrations/rt-qwen2.5-3b-instruct-q4km-4090-ctx8192.json |
Planned, not measured
| Calibration | Plan | Model | Hardware | Quantization | Status | Note |
|---|---|---|---|---|---|---|
planned-qwen2.5-7b-instruct-q4km-4090 | P-01 | Qwen2.5-7B-Instruct | GeForce RTX 4090 24GB (desktop) | gguf-q4_k_m | Planned | BENCHMARKS.md Planned Run Plan P-01. Not a measured calibration. |
planned-qwen2.5-32b-instruct-q4km-4090 | P-02 | Qwen2.5-32B-Instruct | GeForce RTX 4090 24GB (desktop) | gguf-q4_k_m | Planned | BENCHMARKS.md Planned Run Plan P-02. Not a measured calibration. |
planned-dgx-spark-unified-memory | P-03 | — | NVIDIA DGX Spark (GB10, 128 GB unified) | — | Planned | BENCHMARKS.md Planned Run Plan P-03. Not a measured calibration. |
Community recipe benchmarks
Each row below is a community recipe benchmark copied from the recipe file. It is not a lab calibration. Its speed figure is not in the calibration table.
| Status | Calibration | Metric | Value | Unit | Source |
|---|---|---|---|---|---|
| Community recipe benchmark | deepseek-v4.1-flash-2x-spark-exl3 | decode-tok-s | 31.6 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-2x-spark-exl3 | decode-tok-s | 42.5 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-2x-spark-exl3 | decode-tok-s | 42.8 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-2x-spark-exl3 | prefill-tok-s | 1,054.7 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-2x-spark-exl3 | prefill-tok-s | 872.6 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-3x-spark-sglang | decode-tok-s | 37.9 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-3x-spark-sglang | decode-tok-s | 58.9 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-3x-spark-sglang | decode-tok-s | 71.2 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-3x-spark-sglang | decode-tok-s | 78.6 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-4x-spark-full-context | decode-tok-s | 45.4 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-4x-spark-full-context | decode-tok-s | 72.9 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-4x-spark-full-context | decode-tok-s | 103.1 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-4x-spark-full-context | decode-tok-s | 114.1 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-4x-spark-full-context | decode-tok-s | 134.2 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-4x-spark-full-context | prefill-tok-s | 3,350 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-4x-spark-full-context | prefill-tok-s | 3,782 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-4x-spark-full-context | prefill-tok-s | 3,768 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-4x-spark-full-context | prefill-tok-s | 3,531 | tok/s | Source |
| Community recipe benchmark | deepseek-v4.1-flash-4x-spark-full-context | prefill-tok-s | 3,251 | tok/s | Source |
| Community recipe benchmark | deepseek-v4-flash-0731-2x-spark-dspark | decode-tok-s | 75.4 | tok/s | Source |
| Community recipe benchmark | deepseek-v4-flash-0731-2x-spark-dspark | decode-tok-s | 104.9 | tok/s | Source |
| Community recipe benchmark | deepseek-v4-flash-0731-2x-spark-dspark | decode-tok-s | 164.5 | tok/s | Source |
| Community recipe benchmark | deepseek-v4-flash-0731-2x-spark-dspark | decode-tok-s | 191.2 | tok/s | Source |
| Community recipe benchmark | qwen3.8-27b-rtx5090-32gb-nvfp4 | decode-tok-s | 160 | tok/s | Source |