# DeepSeek V4.1: four-Spark boot-10 baseline A clearly identified TP4 baseline using disk-backed Engram and DSpark. Read the boot-10 section separately from later EXL3 and SGLang lanes. - Type: Performance report - Evidence: source_reported; site reproduced: false - Model family: DeepSeek - Model: DeepSeek-V4.1-Flash release checkpoint - Original author: tonyd2wild - Original project: https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark - Pinned document: https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/blob/d45538f67366c62da668c35ec1afd4fcd0e4c8b4/README.md - Source revision: d45538f67366c62da668c35ec1afd4fcd0e4c8b4 - Source commit time (not experiment date): 2026-09-19T13:16:08Z - Checked: 2026-10-17 - Experiment date: None - Runtime topics: vLLM - Hardware topics: spark, nvidia - Reported setup: {"system": "4 x NVIDIA DGX Spark", "node_count": 4, "parallelism": "TP4", "kv_dtype": null, "context_limit_tokens": 300000, "experiment_date": null} - Performance figures: attributed selections with per-point conditions; no cross-source ranking ## Conditions Unknown is null. Recipe options are not measured benchmark settings. ```json { "conditions": { "model_revision": null, "weight_hash": null, "runtime_execution_revision": null, "input_tokens": null, "output_tokens": null, "concurrency": null, "cache_state": null, "quantization": null, "speculative_decoding": null, "temperature": null, "top_p": null }, "test_profiles": [ { "label": "Boot 10 streaming sweep", "evidence_url": "https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/blob/d45538f67366c62da668c35ec1afd4fcd0e4c8b4/README.md#benchmark-boot-10-the-serving-config", "conditions": { "model_revision": null, "weight_hash": null, "runtime_execution_revision": null, "input_tokens": "approximately 30–120", "output_tokens": "budget 150–256; natural stopping", "concurrency": null, "cache_state": "after a warmup; see per-row prefill policy", "quantization": "MXFP4 experts / MXFP8 dense", "speculative_decoding": "DSpark k=5", "temperature": 0, "top_p": null, "thinking": false } } ], "recipe_parameters": null, "deployment_options": [] } ``` ## Reading notes - The selected baseline uses four Sparks, TP4 and DSpark k=5. - Its short-prompt streaming sweep retains thinking and sampling controls. ## Limits - Later EXL3, TP3 and 1M-context boots are separate experiments. - No 1017 checkpoint performance is established. ## License declaration {"declared_spdx": "MIT", "scope": "Repository metadata only; per-file, model, data and conversion terms remain separate."} ## Editorial answer Boot 10 uses four Sparks at TP4 with node-local disk-backed Engram rows, DSpark k5, tools and vision enabled, 300K configured context and GPU memory utilization 0.80. Later EXL3, TP3 and longer-context boots in the same repository are separate experiments. This page keeps the boot-10 measurements anchored to their own section. By Run AI Here editorial · 2026-10-17 AI-assisted source reading and original editorial synthesis; no independent model execution. ## Which boot does this article cover? Boot 10 uses four Sparks at TP4 with node-local disk-backed Engram rows, DSpark k5, tools and vision enabled, 300K configured context and GPU memory utilization 0.80. Later EXL3, TP3 and longer-context boots in the same repository are separate experiments. This page keeps the boot-10 measurements anchored to their own section. Sources: https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/blob/d45538f67366c62da668c35ec1afd4fcd0e4c8b4/README.md#L113 ## Why are aggregate and decode rates different? The author counts tokens from the server usage field rather than stream chunks. Aggregate output divides all stream tokens by batch wall time, including the wait for the first token. Per-stream decode counts tokens after the first over time after that first token. At C1 the category means are consequently 37.95 aggregate and 43.12 decode tokens/s, with 0.441-second mean TTFT. Sources: https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/blob/d45538f67366c62da668c35ec1afd4fcd0e4c8b4/README.md#L121 ## How repeatable is the published sweep? The fixed category prompt set is warmed, sampled at temperature zero with thinking off, and released in concurrent batches. Each cell has one batch. The report warns of a GB10 slow-clock state that can move an individual cell substantially. Its counting ceiling is excluded from the eight-category mean, and should stay excluded when quoting that mean. Sources: https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/blob/d45538f67366c62da668c35ec1afd4fcd0e4c8b4/README.md#L124 ## What can the data be used for? Use the report to understand this TP4 configuration and its content-dependent latency. It does not establish 1017 performance, a broad workload average or a controlled comparison with later engine lanes. The prefill table uses cold unique prompts and one-token replies, a different profile from the short streaming sweep. Sources: https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/blob/d45538f67366c62da668c35ec1afd4fcd0e4c8b4/README.md#L179 ## Selected source-reported figures Each point retains its own configuration, measurement unit and source. No site reproduction or cross-source ranking. ```json [ { "label": "Eight-category mean, C1", "metric": "aggregate_output", "value": 37.95, "unit": "tokens/s", "statistic": "mean across eight categories; counting excluded", "uncertainty": null, "conditions": { "system": "4 x NVIDIA DGX Spark", "node_count": 4, "parallelism": "TP4", "kv_dtype": null, "context_limit_tokens": 300000, "experiment_date": null, "model": "DeepSeek V4.1 Flash launch checkpoint, not 1017", "speculative_decoding": "DSpark k5", "temperature": 0, "thinking": false, "warmup": "performed", "repetitions": "one batch per cell", "input_tokens": "about 30–120; summary prompt longer", "output_token_budget": "150–256; natural stopping", "weight_precision": "MXFP4 experts / MXFP8 dense", "concurrency": 1 }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/blob/d45538f67366c62da668c35ec1afd4fcd0e4c8b4/README.md#L130", "source_document_sha256": "8c72dbb42a49a0f9b5b5f8a70055a467823993012bdb47cad748b5b0e04e36a7" }, { "label": "Eight-category mean, C1", "metric": "per_stream_decode", "value": 43.12, "unit": "tokens/s", "statistic": "mean across eight categories; counting excluded", "uncertainty": null, "conditions": { "system": "4 x NVIDIA DGX Spark", "node_count": 4, "parallelism": "TP4", "kv_dtype": null, "context_limit_tokens": 300000, "experiment_date": null, "model": "DeepSeek V4.1 Flash launch checkpoint, not 1017", "speculative_decoding": "DSpark k5", "temperature": 0, "thinking": false, "warmup": "performed", "repetitions": "one batch per cell", "input_tokens": "about 30–120; summary prompt longer", "output_token_budget": "150–256; natural stopping", "weight_precision": "MXFP4 experts / MXFP8 dense", "concurrency": 1, "denominator": "after first token" }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/blob/d45538f67366c62da668c35ec1afd4fcd0e4c8b4/README.md#L130", "source_document_sha256": "8c72dbb42a49a0f9b5b5f8a70055a467823993012bdb47cad748b5b0e04e36a7" }, { "label": "Eight-category mean, C1", "metric": "ttft", "value": 0.441, "unit": "seconds", "statistic": "mean across eight categories", "uncertainty": null, "conditions": { "system": "4 x NVIDIA DGX Spark", "node_count": 4, "parallelism": "TP4", "kv_dtype": null, "context_limit_tokens": 300000, "experiment_date": null, "model": "DeepSeek V4.1 Flash launch checkpoint, not 1017", "speculative_decoding": "DSpark k5", "temperature": 0, "thinking": false, "warmup": "performed", "repetitions": "one batch per cell", "input_tokens": "about 30–120; summary prompt longer", "output_token_budget": "150–256; natural stopping", "weight_precision": "MXFP4 experts / MXFP8 dense", "concurrency": 1 }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/blob/d45538f67366c62da668c35ec1afd4fcd0e4c8b4/README.md#L130", "source_document_sha256": "8c72dbb42a49a0f9b5b5f8a70055a467823993012bdb47cad748b5b0e04e36a7" }, { "label": "Eight-category mean, C6", "metric": "aggregate_output", "value": 131.86, "unit": "tokens/s", "statistic": "mean across eight categories; counting excluded", "uncertainty": null, "conditions": { "system": "4 x NVIDIA DGX Spark", "node_count": 4, "parallelism": "TP4", "kv_dtype": null, "context_limit_tokens": 300000, "experiment_date": null, "model": "DeepSeek V4.1 Flash launch checkpoint, not 1017", "speculative_decoding": "DSpark k5", "temperature": 0, "thinking": false, "warmup": "performed", "repetitions": "one batch per cell", "input_tokens": "about 30–120; summary prompt longer", "output_token_budget": "150–256; natural stopping", "weight_precision": "MXFP4 experts / MXFP8 dense", "concurrency": 6 }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark/blob/d45538f67366c62da668c35ec1afd4fcd0e4c8b4/README.md#L135", "source_document_sha256": "8c72dbb42a49a0f9b5b5f8a70055a467823993012bdb47cad748b5b0e04e36a7" } ] ```