Recipes, with the numbers they produced.
Every recipe was built and measured on the hardware it names. Commands, versions, throughput, and the failure modes — including the experiments that were measured and not adopted.
Production recipes
6 published| No. | Recipe | Runtime | Decode | Prefill | Context | Measured | |
|---|---|---|---|---|---|---|---|
| 06 | DeepSeek-V4.1-Flash-1017 on 4× DGX Spark with SGLangDeepSeek-V4.1-Flash · 4× DGX Spark GB10 · TP4/EP4 | SGLang 1.7.1 | 66.3 | — | — | 2026-10-18 | → |
| 05 | DeepSeek-V4.1-Flash on 4× DGX Spark with SGLangDeepSeek-V4.1-Flash · 4× DGX Spark GB10 · TP4/EP4 | SGLang 1.7.1 | 94.1 | 4,113 | 524K | 2026-09-17 | → |
| 04 | DeepSeek-V4.1-Flash on 4× DGX SparkDeepSeek-V4.1-Flash · 4× DGX Spark GB10 · TP4/EP4 | vLLM 1.1 | 60.8 | 2,143 | 500K | 2026-09-11 | → |
| 03 | DeepSeek-V4-Flash-Vision on 4× DGX SparkDeepSeek-V4-Flash · 4× DGX Spark GB10 · TP4/EP4 | vLLM 3.73-vision | 67.9 | 2,380 | 393K | 2026-09-03 | → |
| 02 | GLM-5.3-Flash on 4× DGX SparkGLM-5.3-Flash · 4× DGX Spark GB10 · TP4/EP4 | vLLM 1.2 | 89.7 | 2,070 | 393K | 2026-08-31 | → |
| 01 | DeepSeek-V4-Flash on 4× DGX SparkDeepSeek-V4-Flash · 4× DGX Spark GB10 · TP4/EP4 | vLLM 3.76.1 | 71 | 2,200 | 393K | 2026-08-20 | → |
Decode is single-stream at temperature 0 with completion_tokens accounting; prefill is cold and unique-content at 32K–36K depth. A dash means that run did not re-measure the figure. Each recipe states its own prompt set, so treat rows as indicative across recipes and exact within one.
Runtime guides
3 guides| No. | Guide | Hardware | Runtime | Precision | |
|---|---|---|---|---|---|
| 03 | GLM-5.2 Hybrid FP8 / NVFP4 / MXFP4GLM-5.2 · vLLM · multi-node | 4× DGX Spark GB10 (121GB each) | vLLM | Hybrid FP8 / NVFP4 / MXFP4 | → |
| 02 | GLM-5.3-Flash TP4GLM-5.3-Flash · vLLM · multi-node | 4× DGX Spark GB10 (121GB each) | vLLM | FP8 e4m3 (or NVFP4) | → |
| 01 | DiffusionGemma 26B-A4B-it INT8 DynamicDiffusionGemma · vLLM · single-GPU | 2× RTX 3090 (or 1× GPU ≥32GB VRAM) | vLLM | INT8 W8A8 dynamic | → |
Runtime guides are deployments written up from community threads and model cards, with the reported results attributed. Where a claim was not reproduced here, it says so.