THE SOURCE IN CONTEXT
The v2.4.0 promotion arm serves the pinned GDN-MSE checkpoint, with an extra MXFP8 prefill shard. NVFP4 expert weights, MXFP8 paths and a refitted four-draft MTP head are separate components. The experimental image tag is retained; the later release-image digest does not prove the exact cold-test image digest. Repository release names are specific to this one-Spark recipe.
Which quantization and build were tested?
The v2.4.0 promotion arm serves the pinned GDN-MSE checkpoint, with an extra MXFP8 prefill shard. NVFP4 expert weights, MXFP8 paths and a refitted four-draft MTP head are separate components. The experimental image tag is retained; the later release-image digest does not prove the exact cold-test image digest. Repository release names are specific to this one-Spark recipe.
Source: ursuciprian · pinned document ursuciprian · pinned document
What does the long-prompt curve measure?
The selected cold-prefix observations cover 16384, 65536 and 131072 input tokens at C1, temperature zero and one requested output token. Values are means of four boots with three cold prompts each; reported SD is between boot means. A cold prefix does not establish cold operating-system page cache. These are prefill/TTFT observations, not long-response decode or broad reasoning accuracy.
Source: ursuciprian · pinned document ursuciprian · pinned document
How should memory limits be read?
The recipe allocates a 14 GiB FP8 KV pool and reports 993754 token capacity, while its request window is 262144 and scheduler ceiling is eight requests. The token pool is shared: these numbers do not demonstrate eight simultaneous full windows. PLE uses NVMe mmap and page cache. The 0.01 GPU-memory-utilization setting is a startup guard with explicit KV bytes, not measured one-percent memory use.
Source: ursuciprian · pinned document ursuciprian · pinned document
How should two-Spark comparisons be selected?
The project numbers releases independently for one-Spark and two-Spark setups. Its capability table also combines dated measurements when a newer build has no corresponding cell. Keep setup, release, workload and original run path together instead of treating every headline cell as one matched scaling experiment.
Source: ursuciprian · pinned document
Reported figures and their conditions
Selected facts reported by the original author. Each row has its own configuration and measurement boundary. Run AI Here has not reproduced these tests.
| Configuration / workload | Metric | Reported value | Evidence |
|---|---|---|---|
| 1x v2.4.0 16384-token ttft | ttft | 7.104 ± 0.016 SD smean of 4 boots x 3 cold prompts, sd between boots | Original row ↗ |
| 1x v2.4.0 16384-token prefill_rate | prefill rate | 2306.4 ± 5.2 SD tokens/smean of 4 boots x 3 cold prompts, sd between boots | Original row ↗ |
| 1x v2.4.0 65536-token ttft | ttft | 30.723 ± 0.06 SD smean of 4 boots x 3 cold prompts, sd between boots | Original row ↗ |
| 1x v2.4.0 65536-token prefill_rate | prefill rate | 2133.2 ± 4.2 SD tokens/smean of 4 boots x 3 cold prompts, sd between boots | Original row ↗ |
| 1x v2.4.0 131072-token ttft | ttft | 70.0 ± 0.137 SD smean of 4 boots x 3 cold prompts, sd between boots | Original row ↗ |
| 1x v2.4.0 131072-token prefill_rate | prefill rate | 1872.5 ± 3.7 SD tokens/smean of 4 boots x 3 cold prompts, sd between boots | Original row ↗ |
| 1x v2.4.0 allocated KV pool | kv pool tokens | 993754 tokensboot allocation observed on every reported boot | Original row ↗ |
1x v2.4.0 16384-token ttft · ttft · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | ursuciprian/Qwen3.8-Flash-Next-NVFP4-GDN-MSE |
| model_revision | 03f4a0570496bbe741189f4a2a7d611679e85406 |
| extra_prefill_artifact | local-inference-lab/Qwen3.8-Flash-Next-NVFP4 |
| extra_prefill_revision | 7c4f1bc1a2d6847e0cbc01ac6b823f00251de8dd |
| weight_precision | NVFP4 experts; MXFP8 dense/attention; GDN NVFP4 decode and MXFP8 prefill |
| system | one DGX Spark GB10 with 128 GB unified memory, local NVMe |
| node_count | 1 |
| parallelism | TP1 |
| engine | vLLM + b12x |
| engine_version | Not specified |
| runtime_execution_revision | Not specified |
| runtime_revision_prefix | 10d0acc8 files over 5dad364d |
| b12x_revision_prefix | a1ad1b69 |
| executed_image_tag | ghcr.io/ursuciprian/spark-vllm-b12x:k103-20261015-a1ad1b69-10d0acc8 |
| executed_image_digest | Not specified |
| release_image_digest | sha256:30fe678ed7ae5d4deb277be137497bcf597e46e829471b7bed24edd4294a84f1 |
| release_image_is_exact_cold_test_image | False |
| recipe_release | 1x v2.4.0 promotion A/B; k103 cb test arm, not the release-image ship check |
| experiment_date | 2026-10-16 |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_tokens | 4 |
| kv_dtype | fp8 |
| GPU_memory_utilization_startup_guard | 0.01 |
| configured_KV_pool_bytes | 15032385536 |
| PLE_backing | NVMe mmap, page cache; reported 26.8 GiB PLE table |
| concurrency | 1 |
| cache_state | cold prefix, no cache hit; not a claimed cold OS page cache |
| temperature | 0 |
| thinking | Not specified |
| output_tokens | 1 |
| actual_output_tokens_per_request | Not specified |
| GPU_clock_cap_MHz | Not specified |
| input_tokens | 16384 |
| repetitions | 4 boots x 3 cold prompts; SD between boots |
| harness | cold.py (k101) |
| source_run_directory | 1x:results/k103-1x-prefill-combo-ab-20261016-1206/cb-01/ |
| source_metric | ttft_s |
| workload | long-prompt |
{'kind': 'SD between boot means, not confidence interval', 'standard_deviation': 0.016}
1x v2.4.0 16384-token prefill_rate · prefill rate · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | ursuciprian/Qwen3.8-Flash-Next-NVFP4-GDN-MSE |
| model_revision | 03f4a0570496bbe741189f4a2a7d611679e85406 |
| extra_prefill_artifact | local-inference-lab/Qwen3.8-Flash-Next-NVFP4 |
| extra_prefill_revision | 7c4f1bc1a2d6847e0cbc01ac6b823f00251de8dd |
| weight_precision | NVFP4 experts; MXFP8 dense/attention; GDN NVFP4 decode and MXFP8 prefill |
| system | one DGX Spark GB10 with 128 GB unified memory, local NVMe |
| node_count | 1 |
| parallelism | TP1 |
| engine | vLLM + b12x |
| engine_version | Not specified |
| runtime_execution_revision | Not specified |
| runtime_revision_prefix | 10d0acc8 files over 5dad364d |
| b12x_revision_prefix | a1ad1b69 |
| executed_image_tag | ghcr.io/ursuciprian/spark-vllm-b12x:k103-20261015-a1ad1b69-10d0acc8 |
| executed_image_digest | Not specified |
| release_image_digest | sha256:30fe678ed7ae5d4deb277be137497bcf597e46e829471b7bed24edd4294a84f1 |
| release_image_is_exact_cold_test_image | False |
| recipe_release | 1x v2.4.0 promotion A/B; k103 cb test arm, not the release-image ship check |
| experiment_date | 2026-10-16 |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_tokens | 4 |
| kv_dtype | fp8 |
| GPU_memory_utilization_startup_guard | 0.01 |
| configured_KV_pool_bytes | 15032385536 |
| PLE_backing | NVMe mmap, page cache; reported 26.8 GiB PLE table |
| concurrency | 1 |
| cache_state | cold prefix, no cache hit; not a claimed cold OS page cache |
| temperature | 0 |
| thinking | Not specified |
| output_tokens | 1 |
| actual_output_tokens_per_request | Not specified |
| GPU_clock_cap_MHz | Not specified |
| input_tokens | 16384 |
| repetitions | 4 boots x 3 cold prompts; SD between boots |
| harness | cold.py (k101) |
| source_run_directory | 1x:results/k103-1x-prefill-combo-ab-20261016-1206/cb-01/ |
| source_metric | pp |
| workload | long-prompt |
{'kind': 'SD between boot means, not confidence interval', 'standard_deviation': 5.2}
1x v2.4.0 65536-token ttft · ttft · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | ursuciprian/Qwen3.8-Flash-Next-NVFP4-GDN-MSE |
| model_revision | 03f4a0570496bbe741189f4a2a7d611679e85406 |
| extra_prefill_artifact | local-inference-lab/Qwen3.8-Flash-Next-NVFP4 |
| extra_prefill_revision | 7c4f1bc1a2d6847e0cbc01ac6b823f00251de8dd |
| weight_precision | NVFP4 experts; MXFP8 dense/attention; GDN NVFP4 decode and MXFP8 prefill |
| system | one DGX Spark GB10 with 128 GB unified memory, local NVMe |
| node_count | 1 |
| parallelism | TP1 |
| engine | vLLM + b12x |
| engine_version | Not specified |
| runtime_execution_revision | Not specified |
| runtime_revision_prefix | 10d0acc8 files over 5dad364d |
| b12x_revision_prefix | a1ad1b69 |
| executed_image_tag | ghcr.io/ursuciprian/spark-vllm-b12x:k103-20261015-a1ad1b69-10d0acc8 |
| executed_image_digest | Not specified |
| release_image_digest | sha256:30fe678ed7ae5d4deb277be137497bcf597e46e829471b7bed24edd4294a84f1 |
| release_image_is_exact_cold_test_image | False |
| recipe_release | 1x v2.4.0 promotion A/B; k103 cb test arm, not the release-image ship check |
| experiment_date | 2026-10-16 |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_tokens | 4 |
| kv_dtype | fp8 |
| GPU_memory_utilization_startup_guard | 0.01 |
| configured_KV_pool_bytes | 15032385536 |
| PLE_backing | NVMe mmap, page cache; reported 26.8 GiB PLE table |
| concurrency | 1 |
| cache_state | cold prefix, no cache hit; not a claimed cold OS page cache |
| temperature | 0 |
| thinking | Not specified |
| output_tokens | 1 |
| actual_output_tokens_per_request | Not specified |
| GPU_clock_cap_MHz | Not specified |
| input_tokens | 65536 |
| repetitions | 4 boots x 3 cold prompts; SD between boots |
| harness | cold.py (k101) |
| source_run_directory | 1x:results/k103-1x-prefill-combo-ab-20261016-1206/cb-01/ |
| source_metric | ttft_s |
| workload | long-prompt |
{'kind': 'SD between boot means, not confidence interval', 'standard_deviation': 0.06}
1x v2.4.0 65536-token prefill_rate · prefill rate · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | ursuciprian/Qwen3.8-Flash-Next-NVFP4-GDN-MSE |
| model_revision | 03f4a0570496bbe741189f4a2a7d611679e85406 |
| extra_prefill_artifact | local-inference-lab/Qwen3.8-Flash-Next-NVFP4 |
| extra_prefill_revision | 7c4f1bc1a2d6847e0cbc01ac6b823f00251de8dd |
| weight_precision | NVFP4 experts; MXFP8 dense/attention; GDN NVFP4 decode and MXFP8 prefill |
| system | one DGX Spark GB10 with 128 GB unified memory, local NVMe |
| node_count | 1 |
| parallelism | TP1 |
| engine | vLLM + b12x |
| engine_version | Not specified |
| runtime_execution_revision | Not specified |
| runtime_revision_prefix | 10d0acc8 files over 5dad364d |
| b12x_revision_prefix | a1ad1b69 |
| executed_image_tag | ghcr.io/ursuciprian/spark-vllm-b12x:k103-20261015-a1ad1b69-10d0acc8 |
| executed_image_digest | Not specified |
| release_image_digest | sha256:30fe678ed7ae5d4deb277be137497bcf597e46e829471b7bed24edd4294a84f1 |
| release_image_is_exact_cold_test_image | False |
| recipe_release | 1x v2.4.0 promotion A/B; k103 cb test arm, not the release-image ship check |
| experiment_date | 2026-10-16 |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_tokens | 4 |
| kv_dtype | fp8 |
| GPU_memory_utilization_startup_guard | 0.01 |
| configured_KV_pool_bytes | 15032385536 |
| PLE_backing | NVMe mmap, page cache; reported 26.8 GiB PLE table |
| concurrency | 1 |
| cache_state | cold prefix, no cache hit; not a claimed cold OS page cache |
| temperature | 0 |
| thinking | Not specified |
| output_tokens | 1 |
| actual_output_tokens_per_request | Not specified |
| GPU_clock_cap_MHz | Not specified |
| input_tokens | 65536 |
| repetitions | 4 boots x 3 cold prompts; SD between boots |
| harness | cold.py (k101) |
| source_run_directory | 1x:results/k103-1x-prefill-combo-ab-20261016-1206/cb-01/ |
| source_metric | pp |
| workload | long-prompt |
{'kind': 'SD between boot means, not confidence interval', 'standard_deviation': 4.2}
1x v2.4.0 131072-token ttft · ttft · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | ursuciprian/Qwen3.8-Flash-Next-NVFP4-GDN-MSE |
| model_revision | 03f4a0570496bbe741189f4a2a7d611679e85406 |
| extra_prefill_artifact | local-inference-lab/Qwen3.8-Flash-Next-NVFP4 |
| extra_prefill_revision | 7c4f1bc1a2d6847e0cbc01ac6b823f00251de8dd |
| weight_precision | NVFP4 experts; MXFP8 dense/attention; GDN NVFP4 decode and MXFP8 prefill |
| system | one DGX Spark GB10 with 128 GB unified memory, local NVMe |
| node_count | 1 |
| parallelism | TP1 |
| engine | vLLM + b12x |
| engine_version | Not specified |
| runtime_execution_revision | Not specified |
| runtime_revision_prefix | 10d0acc8 files over 5dad364d |
| b12x_revision_prefix | a1ad1b69 |
| executed_image_tag | ghcr.io/ursuciprian/spark-vllm-b12x:k103-20261015-a1ad1b69-10d0acc8 |
| executed_image_digest | Not specified |
| release_image_digest | sha256:30fe678ed7ae5d4deb277be137497bcf597e46e829471b7bed24edd4294a84f1 |
| release_image_is_exact_cold_test_image | False |
| recipe_release | 1x v2.4.0 promotion A/B; k103 cb test arm, not the release-image ship check |
| experiment_date | 2026-10-16 |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_tokens | 4 |
| kv_dtype | fp8 |
| GPU_memory_utilization_startup_guard | 0.01 |
| configured_KV_pool_bytes | 15032385536 |
| PLE_backing | NVMe mmap, page cache; reported 26.8 GiB PLE table |
| concurrency | 1 |
| cache_state | cold prefix, no cache hit; not a claimed cold OS page cache |
| temperature | 0 |
| thinking | Not specified |
| output_tokens | 1 |
| actual_output_tokens_per_request | Not specified |
| GPU_clock_cap_MHz | Not specified |
| input_tokens | 131072 |
| repetitions | 4 boots x 3 cold prompts; SD between boots |
| harness | cold.py (k101) |
| source_run_directory | 1x:results/k103-1x-prefill-combo-ab-20261016-1206/cb-01/ |
| source_metric | ttft_s |
| workload | long-prompt |
{'kind': 'SD between boot means, not confidence interval', 'standard_deviation': 0.137}
1x v2.4.0 131072-token prefill_rate · prefill rate · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | ursuciprian/Qwen3.8-Flash-Next-NVFP4-GDN-MSE |
| model_revision | 03f4a0570496bbe741189f4a2a7d611679e85406 |
| extra_prefill_artifact | local-inference-lab/Qwen3.8-Flash-Next-NVFP4 |
| extra_prefill_revision | 7c4f1bc1a2d6847e0cbc01ac6b823f00251de8dd |
| weight_precision | NVFP4 experts; MXFP8 dense/attention; GDN NVFP4 decode and MXFP8 prefill |
| system | one DGX Spark GB10 with 128 GB unified memory, local NVMe |
| node_count | 1 |
| parallelism | TP1 |
| engine | vLLM + b12x |
| engine_version | Not specified |
| runtime_execution_revision | Not specified |
| runtime_revision_prefix | 10d0acc8 files over 5dad364d |
| b12x_revision_prefix | a1ad1b69 |
| executed_image_tag | ghcr.io/ursuciprian/spark-vllm-b12x:k103-20261015-a1ad1b69-10d0acc8 |
| executed_image_digest | Not specified |
| release_image_digest | sha256:30fe678ed7ae5d4deb277be137497bcf597e46e829471b7bed24edd4294a84f1 |
| release_image_is_exact_cold_test_image | False |
| recipe_release | 1x v2.4.0 promotion A/B; k103 cb test arm, not the release-image ship check |
| experiment_date | 2026-10-16 |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_tokens | 4 |
| kv_dtype | fp8 |
| GPU_memory_utilization_startup_guard | 0.01 |
| configured_KV_pool_bytes | 15032385536 |
| PLE_backing | NVMe mmap, page cache; reported 26.8 GiB PLE table |
| concurrency | 1 |
| cache_state | cold prefix, no cache hit; not a claimed cold OS page cache |
| temperature | 0 |
| thinking | Not specified |
| output_tokens | 1 |
| actual_output_tokens_per_request | Not specified |
| GPU_clock_cap_MHz | Not specified |
| input_tokens | 131072 |
| repetitions | 4 boots x 3 cold prompts; SD between boots |
| harness | cold.py (k101) |
| source_run_directory | 1x:results/k103-1x-prefill-combo-ab-20261016-1206/cb-01/ |
| source_metric | pp |
| workload | long-prompt |
{'kind': 'SD between boot means, not confidence interval', 'standard_deviation': 3.7}
1x v2.4.0 allocated KV pool · kv pool tokens · conditions
| model | Qwen3.8-Flash-Next |
|---|---|
| model_artifact | ursuciprian/Qwen3.8-Flash-Next-NVFP4-GDN-MSE |
| model_revision | 03f4a0570496bbe741189f4a2a7d611679e85406 |
| extra_prefill_artifact | local-inference-lab/Qwen3.8-Flash-Next-NVFP4 |
| extra_prefill_revision | 7c4f1bc1a2d6847e0cbc01ac6b823f00251de8dd |
| weight_precision | NVFP4 experts; MXFP8 dense/attention; GDN NVFP4 decode and MXFP8 prefill |
| system | one DGX Spark GB10 with 128 GB unified memory, local NVMe |
| node_count | 1 |
| parallelism | TP1 |
| engine | vLLM + b12x |
| engine_version | Not specified |
| runtime_execution_revision | Not specified |
| runtime_revision_prefix | 10d0acc8 files over 5dad364d |
| b12x_revision_prefix | a1ad1b69 |
| executed_image_tag | ghcr.io/ursuciprian/spark-vllm-b12x:k103-20261015-a1ad1b69-10d0acc8 |
| executed_image_digest | Not specified |
| release_image_digest | sha256:30fe678ed7ae5d4deb277be137497bcf597e46e829471b7bed24edd4294a84f1 |
| release_image_is_exact_cold_test_image | False |
| recipe_release | 1x v2.4.0 promotion A/B; k103 cb test arm, not the release-image ship check |
| experiment_date | 2026-10-16 |
| context_limit_tokens | 262144 |
| configured_max_requests | 8 |
| draft_tokens | 4 |
| kv_dtype | fp8 |
| GPU_memory_utilization_startup_guard | 0.01 |
| configured_KV_pool_bytes | 15032385536 |
| PLE_backing | NVMe mmap, page cache; reported 26.8 GiB PLE table |
| concurrency | Not specified |
| cache_state | cold prefix, no cache hit; not a claimed cold OS page cache |
| temperature | Not specified |
| thinking | Not specified |
| output_tokens | Not specified |
| actual_output_tokens_per_request | Not specified |
| GPU_clock_cap_MHz | Not specified |
| input_tokens | Not specified |
| memory_kind | allocated_KV_token_pool_not_measured_peak_RAM |
| max_context_simultaneously_validated | False |
Original source-grounded editorial synthesis with local Spark Qwen assistance and Codex review; no site model benchmark.