# llama.cpp on Spark: Q4_K_M prefill and decode Original Spark report separating prompt processing from single-batch generation at different context depths. - Type: Performance report - Evidence: source_reported; site reproduced: false - Model family: Multiple - Model: Qwen3 / Ministral 3 - Original author: DandinPower - Original project: https://github.com/DandinPower/llama.cpp_bench - Pinned document: https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md - Source revision: a1e9514561e9be68982e1905f48808143082013a - Source commit time (not experiment date): 2025-12-05T13:40:35Z - Checked: 2026-10-17 - Experiment date: None - Runtime topics: llama.cpp - Hardware topics: spark, nvidia - Reported setup: {"system": "NVIDIA DGX Spark / GB10 / 128 GB LPDDR5x", "node_count": 1, "parallelism": null, "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": null, "weight_hash": null, "runtime_execution_revision": null, "input_tokens": null, "output_tokens": null, "concurrency": null, "cache_state": null, "quantization": "Q4_K_M", "speculative_decoding": null, "temperature": null, "top_p": null }, "test_profiles": [ { "label": "Prompt processing", "evidence_url": "https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md", "conditions": { "model_revision": null, "weight_hash": null, "runtime_execution_revision": null, "input_tokens": null, "output_tokens": null, "concurrency": null, "cache_state": null, "quantization": "Q4_K_M", "speculative_decoding": null, "temperature": null, "top_p": null, "context_depth_tokens": [ 512, 16384 ] } }, { "label": "Generation", "evidence_url": "https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md", "conditions": { "model_revision": null, "weight_hash": null, "runtime_execution_revision": null, "input_tokens": null, "output_tokens": null, "concurrency": null, "cache_state": null, "quantization": "Q4_K_M", "speculative_decoding": null, "temperature": null, "top_p": null, "batch_size": 1, "context_depth_tokens": [ 512, 2048 ] } } ], "recipe_parameters": null, "deployment_options": [] } ``` ## Reading notes - Q4_K_M is the selected weight format. - Prefill and decode tables describe different context depths. ## Limits - Exact model and runtime commits are not specified in this report. - A tested context depth is not the machine’s maximum supported context. ## License declaration {"declared_spdx": null, "scope": "Repository metadata only; per-file, model, data and conversion terms remain separate."} ## Editorial answer The report studies Qwen3 and Ministral 3 Q4_K_M artifacts on a single GB10 Spark with 128 GB unified memory. Prompt processing is tabulated at context depths 512 and 16,384; single-batch generation at 512 and 2,048. Context depth is the source table’s label, not a specified generated-output budget or a maximum capacity test. By Run AI Here editorial · 2026-10-17 AI-assisted source reading and original editorial synthesis; no independent model execution. ## What experiment is being described? The report studies Qwen3 and Ministral 3 Q4_K_M artifacts on a single GB10 Spark with 128 GB unified memory. Prompt processing is tabulated at context depths 512 and 16,384; single-batch generation at 512 and 2,048. Context depth is the source table’s label, not a specified generated-output budget or a maximum capacity test. Sources: https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md#L20 ## What is a useful concrete comparison? The selected Qwen3 30B MoE row reports 89.3 decode tokens/s at depth 512 and 83.8 at 2,048, while the 32B dense row reports 10.7 at depth 512. These results illustrate the author’s bandwidth interpretation for this particular pair of Q4_K_M artifacts. They do not establish equal quality or the same outcome for another MoE family. Sources: https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md#L33 ## What prevents a strong reproduction claim? The report does not identify exact weight hashes, executed runtime commit, output lengths, cache policy or repetitions. Those gaps remain explicit. Without them, a later reader cannot know whether a discrepancy comes from kernels, artifacts or workload. A tested context depth also cannot certify the machine’s maximum context. Sources: https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md#L18 ## How is the source reused? This article gives a small attributed factual selection with an original explanation. Repository metadata does not declare a license, so the full result table and prose are left with the original author. No independent run is claimed, and prefill and decode remain separate measurement types. Sources: https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md#L46 ## Selected source-reported figures Each point retains its own configuration, measurement unit and source. No site reproduction or cross-source ranking. ```json [ { "label": "Qwen3 30B MoE, shallow context", "metric": "decode", "value": 89.3, "unit": "tokens/s", "statistic": "reported value", "uncertainty": null, "conditions": { "system": "NVIDIA DGX Spark / GB10 / 128 GB LPDDR5x", "node_count": 1, "parallelism": null, "kv_dtype": null, "context_limit_tokens": null, "experiment_date": null, "runtime": "llama.cpp, executed commit unspecified", "quantization": "Q4_K_M", "batch_size": 1, "input_tokens": null, "output_tokens": null, "repetitions": null, "cache_state": null, "model": "Qwen3 30B MoE", "context_depth_tokens": 512 }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md#L40", "source_document_sha256": "44d415d26c12533d1694fcc7b622f8961ea0a320078177e14c663a5b77028c70" }, { "label": "Qwen3 30B MoE, deeper context", "metric": "decode", "value": 83.8, "unit": "tokens/s", "statistic": "reported value", "uncertainty": null, "conditions": { "system": "NVIDIA DGX Spark / GB10 / 128 GB LPDDR5x", "node_count": 1, "parallelism": null, "kv_dtype": null, "context_limit_tokens": null, "experiment_date": null, "runtime": "llama.cpp, executed commit unspecified", "quantization": "Q4_K_M", "batch_size": 1, "input_tokens": null, "output_tokens": null, "repetitions": null, "cache_state": null, "model": "Qwen3 30B MoE", "context_depth_tokens": 2048 }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md#L40", "source_document_sha256": "44d415d26c12533d1694fcc7b622f8961ea0a320078177e14c663a5b77028c70" }, { "label": "Qwen3 32B dense, shallow context", "metric": "decode", "value": 10.7, "unit": "tokens/s", "statistic": "reported value", "uncertainty": null, "conditions": { "system": "NVIDIA DGX Spark / GB10 / 128 GB LPDDR5x", "node_count": 1, "parallelism": null, "kv_dtype": null, "context_limit_tokens": null, "experiment_date": null, "runtime": "llama.cpp, executed commit unspecified", "quantization": "Q4_K_M", "batch_size": 1, "input_tokens": null, "output_tokens": null, "repetitions": null, "cache_state": null, "model": "Qwen3 32B dense", "context_depth_tokens": 512 }, "evidence_state": "source_reported", "site_reproduced": false, "evidence_url": "https://github.com/DandinPower/llama.cpp_bench/blob/a1e9514561e9be68982e1905f48808143082013a/dgx_spark/report.md#L41", "source_document_sha256": "44d415d26c12533d1694fcc7b622f8961ea0a320078177e14c663a5b77028c70" } ] ```