# DeepSeek V4.1: hybrid RTX/Spark v14 decode Versioned decode campaign comparing one- and two-RTX coordinators with four Spark expert workers, using interleaved sessions. - Type: Performance report - Evidence: source_reported; site reproduced: false - Model family: DeepSeek - Model: DeepSeek-V4.1-Flash @ dba1be0 - Original author: tpurtell - Original project: https://github.com/tpurtell/ds41rt - Pinned document: https://github.com/tpurtell/ds41rt/blob/3067d0684a0b572d88a17d38ca3f7d10599619b4/docs/release-v14-performance.md - Source revision: 3067d0684a0b572d88a17d38ca3f7d10599619b4 - Source commit time (not experiment date): 2026-09-25T11:32:19Z - Checked: 2026-10-17 - Experiment date: None - Runtime topics: DS41RT - Hardware topics: spark, nvidia - Reported setup: {"system": "1 or 2 RTX PRO 6000 Blackwell coordinators plus 4 DGX Spark workers", "node_count": null, "parallelism": "4 Spark expert workers, TP4", "kv_dtype": null, "context_limit_tokens": null, "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": "dba1be0a40aa45a94ad051997016db3960a90277", "weight_hash": null, "runtime_execution_revision": "af9d96923f70af1cd385960e28450e5ec0653ff8", "input_tokens": null, "output_tokens": null, "concurrency": null, "cache_state": null, "quantization": null, "speculative_decoding": null, "temperature": null, "top_p": null }, "test_profiles": [ { "label": "Interleaved v13/v14 decode sessions", "evidence_url": "https://github.com/tpurtell/ds41rt/blob/3067d0684a0b572d88a17d38ca3f7d10599619b4/docs/release-v14-performance.md", "conditions": { "model_revision": "dba1be0a40aa45a94ad051997016db3960a90277", "weight_hash": null, "runtime_execution_revision": "af9d96923f70af1cd385960e28450e5ec0653ff8", "input_tokens": null, "output_tokens": null, "concurrency": "C1–C16 by workload", "cache_state": "discarded warmup battery and sweep; each arm relaunched", "quantization": null, "speculative_decoding": "target-only and local DSpark arms are separate", "temperature": "greedy or 0.7 in separate cases", "top_p": "0.9 for sampled cases", "hardware": "1 or 2 RTX PRO 6000 at 400 W each, standard clocks; 4 DGX Spark", "repetitions": "8 sessions per layout; workload-specific batteries" } } ], "recipe_parameters": null, "deployment_options": [] } ``` ## Reading notes - Each arm is restarted from its own image pair. - Greedy, sampled, concurrent and target-only cases are distinguished. ## Limits - The v14 decode campaign does not refresh historical v6 startup or prefill tables. - A capped RTX PRO 6000 experiment does not demonstrate physical RTX 5090 behavior. ## License declaration {"declared_spdx": "MIT", "scope": "Repository metadata only; per-file, model, data and conversion terms remain separate."} ## Editorial answer Four Spark expert workers are paired with one or two RTX PRO 6000 Blackwell coordinators, each limited to 400 W at standard memory clocks. The report identifies the base checkpoint, runtime source and image IDs. A 400 W RTX PRO 6000 is still that card; the experiment does not establish the behavior of a physical RTX 5090. By Run AI Here editorial · 2026-10-17 AI-assisted source reading and original editorial synthesis; no independent model execution. ## What is this hybrid system? Four Spark expert workers are paired with one or two RTX PRO 6000 Blackwell coordinators, each limited to 400 W at standard memory clocks. The report identifies the base checkpoint, runtime source and image IDs. A 400 W RTX PRO 6000 is still that card; the experiment does not establish the behavior of a physical RTX 5090. Sources: https://github.com/tpurtell/ds41rt/blob/3067d0684a0b572d88a17d38ca3f7d10599619b4/docs/release-v14-performance.md#L6 ## How was v14 compared with v13? Each layout runs eight sessions in an interleaved version order, relaunching the appropriate image pair and discarding warmup. Sessions are the independent units. The published ratios use geometric means, with Welch 95% intervals on session log-ratios. The two-coordinator sampled C1 ratio of 1.377 is therefore a within-campaign version comparison, not a ratio against an unrelated desktop. Sources: https://github.com/tpurtell/ds41rt/blob/3067d0684a0b572d88a17d38ca3f7d10599619b4/docs/release-v14-performance.md#L25 ## Why retain content and traffic type? The release table’s weighted C1 medians are 104.69 tokens/s with one coordinator and 129.67 with two. Counting and code have different rates, and the concurrency sweep uses an aggregate interval from earliest output to final completion. That denominator differs from an individual request’s decode or end-to-end served rate. Sources: https://github.com/tpurtell/ds41rt/blob/3067d0684a0b572d88a17d38ca3f7d10599619b4/docs/release-v14-performance.md#L124 ## What are the limitations and disclosed artifacts? Some greedy outputs differ across arms through batch-shape numerics, although the author reports all objective checks passed. A schema-JSON probe initially hit the prompt cache and was rerun with a fresh nonce; the table uses that rerun. Historical startup and prefill tables belong to older campaigns. Keep the fresh decode evidence separate from those retained history sections. Sources: https://github.com/tpurtell/ds41rt/blob/3067d0684a0b572d88a17d38ca3f7d10599619b4/docs/release-v14-performance.md#L79 ## Selected source-reported figures Each point retains its own configuration, measurement unit and source. No site reproduction or cross-source ranking. ```json [ { "label": "One RTX coordinator + four Sparks, weighted DSpark", "metric": "decode", "value": 104.69, "unit": "tokens/s", "statistic": "median of three repeats", "uncertainty": null, "conditions": { "system": "1 or 2 RTX PRO 6000 Blackwell coordinators plus 4 DGX Spark workers", "node_count": null, "parallelism": "4 Spark expert workers, TP4", "kv_dtype": null, "context_limit_tokens": null, "experiment_date": "2026-09-24", "model_revision": "dba1be0a40aa45a94ad051997016db3960a90277", "runtime_execution_revision": "af9d96923f70af1cd385960e28450e5ec0653ff8", "worker_count": 4, "worker_parallelism": "Spark TP4", "gpu_power_watts_each": 400, "repetitions": 3, "statistic": "median", "concurrency": 1, "coordinator_count": 1 }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/tpurtell/ds41rt/blob/3067d0684a0b572d88a17d38ca3f7d10599619b4/docs/release-v14-performance.md#L105", "source_document_sha256": "b1d31e955106b701569c49476026ad674a838d376f93294395ae0b198d4fd6ee" }, { "label": "Two RTX coordinators + four Sparks, weighted DSpark", "metric": "decode", "value": 129.67, "unit": "tokens/s", "statistic": "median of three repeats", "uncertainty": null, "conditions": { "system": "1 or 2 RTX PRO 6000 Blackwell coordinators plus 4 DGX Spark workers", "node_count": null, "parallelism": "4 Spark expert workers, TP4", "kv_dtype": null, "context_limit_tokens": null, "experiment_date": "2026-09-24", "model_revision": "dba1be0a40aa45a94ad051997016db3960a90277", "runtime_execution_revision": "af9d96923f70af1cd385960e28450e5ec0653ff8", "worker_count": 4, "worker_parallelism": "Spark TP4", "gpu_power_watts_each": 400, "repetitions": 3, "statistic": "median", "concurrency": 1, "coordinator_count": 2 }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/tpurtell/ds41rt/blob/3067d0684a0b572d88a17d38ca3f7d10599619b4/docs/release-v14-performance.md#L105", "source_document_sha256": "b1d31e955106b701569c49476026ad674a838d376f93294395ae0b198d4fd6ee" }, { "label": "One RTX coordinator + four Sparks, target only", "metric": "decode", "value": 59.84, "unit": "tokens/s", "statistic": "median of three repeats", "uncertainty": null, "conditions": { "system": "1 or 2 RTX PRO 6000 Blackwell coordinators plus 4 DGX Spark workers", "node_count": null, "parallelism": "4 Spark expert workers, TP4", "kv_dtype": null, "context_limit_tokens": null, "experiment_date": "2026-09-24", "model_revision": "dba1be0a40aa45a94ad051997016db3960a90277", "runtime_execution_revision": "af9d96923f70af1cd385960e28450e5ec0653ff8", "worker_count": 4, "worker_parallelism": "Spark TP4", "gpu_power_watts_each": 400, "repetitions": 3, "statistic": "median", "concurrency": 1, "coordinator_count": 1, "speculative_decoding": false }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/tpurtell/ds41rt/blob/3067d0684a0b572d88a17d38ca3f7d10599619b4/docs/release-v14-performance.md#L105", "source_document_sha256": "b1d31e955106b701569c49476026ad674a838d376f93294395ae0b198d4fd6ee" }, { "label": "Two RTX coordinators, sampled C1 v14/v13", "metric": "decode_ratio", "value": 1.377, "unit": "ratio", "statistic": "geometric mean of session-score ratio", "uncertainty": { "interval_type": "Welch 95% interval on session log-ratios", "lower": 1.369, "upper": 1.385 }, "conditions": { "system": "1 or 2 RTX PRO 6000 Blackwell coordinators plus 4 DGX Spark workers", "node_count": null, "parallelism": "4 Spark expert workers, TP4", "kv_dtype": null, "context_limit_tokens": null, "experiment_date": "2026-09-24", "model_revision": "dba1be0a40aa45a94ad051997016db3960a90277", "runtime_execution_revision": "af9d96923f70af1cd385960e28450e5ec0653ff8", "worker_count": 4, "worker_parallelism": "Spark TP4", "gpu_power_watts_each": 400, "repetitions": "8 interleaved sessions per layout; four per arm", "statistic": "median", "concurrency": 1, "coordinator_count": 2, "temperature": 0.7, "top_p": 0.9 }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/tpurtell/ds41rt/blob/3067d0684a0b572d88a17d38ca3f7d10599619b4/docs/release-v14-performance.md#L39", "source_document_sha256": "b1d31e955106b701569c49476026ad674a838d376f93294395ae0b198d4fd6ee" } ] ```