THE SOURCE IN CONTEXT
The a14 scheduler change raises cold eight-request extraction aggregate rate from 234.57 to 306.66 tokens/s while all 24 answers pass in each arm. This is long output, not a 128K input test: the single-request fixture log has 1498 input tokens. a13 quality scores and older needle runs keep their original build labels.
What is the eight-node recipe?
The checkpoint is XiaomiMiMo/MiMo-V2.6-Pro-RL at full revision 73875d00b30a89ef8cc353a0b60b0e9f9561952d. Eight GB10 nodes use TP8/EP1, two switched200G RoCEv2 subnets at MTU9000, packed MXFP4 experts on Marlin W4A16, DFlash8 adaptive verify and FA4. The base SGLang SHA and a14 local image ID are pinned separately; a complete patched execution-tree revision and registry digest remain unknown.
What changes in the paired extraction logs?
The a13 scheduler counts a finishing prefill request against slots it already holds, leaving the eighth request queued. Across three cold eight-request waves, a13 has43908 completion tokens/187.18s and a14 has 43900/143.16s. Both pass24/24. Arithmetic mean per-stream rate falls 39.46 to 38.4 while complete-wave aggregate rises 234.57 to 306.66, showing why the denominators must stay separate.
Source: rhys101 · pinned document rhys101 · pinned document rhys101 · pinned document
How long is long extraction?
The C1 source log records 1498 prompt tokens for every fixture, with roughly 1829 output tokens. C8 logs omit prompt counts. Unique salts and prefix-cache flushes make these cold extraction requests, but they are not128K-plus input workloads. Separate a13 cold prefill medians use nominal5K/20K/81K/163K labels.
Source: rhys101 · pinned document rhys101 · pinned document rhys101 · pinned document
Which quality and long-context checks apply to which build?
GSM8K first200 fixed-order questions gives 196 correct on a13, greedy/thinking off with output cap1024 and a custom numeric extractor. No a14 GSM8K rerun is reported. Three-fact needle results at 127803 and 244168 tokens belong to a5/a6, not a13/a14. Finite smoke checks include vision; audio/video and peak runtime RAM are unmeasured.
Source: rhys101 · pinned document rhys101 · pinned document rhys101 · 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 |
|---|---|---|---|
| coding C1 | decode rate | 91.1 tokens/ssource-reported observation; dispersion unreported | Original row ↗ |
| coding C8 per stream | decode rate | 42.3 tokens/ssource-reported observation; dispersion unreported | Original row ↗ |
| eight categories C8 aggregate | aggregate output rate | 190.8 tokens/ssource-reported observation; dispersion unreported | Original row ↗ |
| a13 cold CSV extraction C8 | aggregate output rate | 234.57 tokens/ssource-reported observation; dispersion unreported | Original row ↗ |
| a14 cold CSV extraction C8 | aggregate output rate | 306.66 tokens/ssource-reported observation; dispersion unreported | Original row ↗ |
| a14 cold CSV extraction C1 | aggregate output rate | 104.6 tokens/ssource-reported observation; dispersion unreported | Original row ↗ |
| a13 GSM8K first 200 | accuracy | 98.0 percentsource-reported observation; dispersion unreported | Original row ↗ |
| a13 5K cold prefill | prefill rate | 1880 tokens/smedian of three source runs; no SD | Original row ↗ |
| a13 163K cold prefill | prefill rate | 1499 tokens/smedian of three source runs; no SD | Original row ↗ |
coding C1 · decode rate · conditions
| model | MiMo-V2.6-Pro-RL |
|---|---|
| model_artifact | XiaomiMiMo/MiMo-V2.6-Pro-RL |
| model_revision | 73875d00b30a89ef8cc353a0b60b0e9f9561952d |
| engine | patched SGLang with Marlin W4A16 MXFP4 experts, DFlash8 adaptive verify and FA4 |
| base_engine_revision | 06008c170dae333e22caceb51f36224071f9b5b3 |
| runtime_execution_revision | Not specified |
| executed_image_digest | Not specified |
| executed_image_id | Not specified |
| system | eight DGX Spark GB10, two switched 200G RoCEv2 subnets, MTU9000 |
| node_count | 8 |
| parallelism | TP8/EP1 |
| weight_precision | MXFP4 expert storage; Marlin W4A16; optional FP8 o_proj and draft |
| kv_dtype | Not specified |
| experiment_date | 2026-09-24 |
| measured_build | a13 |
| temperature | 0 |
| thinking | False |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | Not specified |
| cache_state | Not specified |
| peak_runtime_RAM | Not specified |
| context_limit_tokens | Not specified |
| drafter_revision | Not specified |
| concurrency | 1 |
| metric_denominator | source per-stream decode |
coding C8 per stream · decode rate · conditions
| model | MiMo-V2.6-Pro-RL |
|---|---|
| model_artifact | XiaomiMiMo/MiMo-V2.6-Pro-RL |
| model_revision | 73875d00b30a89ef8cc353a0b60b0e9f9561952d |
| engine | patched SGLang with Marlin W4A16 MXFP4 experts, DFlash8 adaptive verify and FA4 |
| base_engine_revision | 06008c170dae333e22caceb51f36224071f9b5b3 |
| runtime_execution_revision | Not specified |
| executed_image_digest | Not specified |
| executed_image_id | Not specified |
| system | eight DGX Spark GB10, two switched 200G RoCEv2 subnets, MTU9000 |
| node_count | 8 |
| parallelism | TP8/EP1 |
| weight_precision | MXFP4 expert storage; Marlin W4A16; optional FP8 o_proj and draft |
| kv_dtype | Not specified |
| experiment_date | 2026-09-24 |
| measured_build | a13 |
| temperature | 0 |
| thinking | False |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | Not specified |
| cache_state | Not specified |
| peak_runtime_RAM | Not specified |
| context_limit_tokens | Not specified |
| drafter_revision | Not specified |
| concurrency | 8 |
| metric_denominator | source per-stream decode |
eight categories C8 aggregate · aggregate output rate · conditions
| model | MiMo-V2.6-Pro-RL |
|---|---|
| model_artifact | XiaomiMiMo/MiMo-V2.6-Pro-RL |
| model_revision | 73875d00b30a89ef8cc353a0b60b0e9f9561952d |
| engine | patched SGLang with Marlin W4A16 MXFP4 experts, DFlash8 adaptive verify and FA4 |
| base_engine_revision | 06008c170dae333e22caceb51f36224071f9b5b3 |
| runtime_execution_revision | Not specified |
| executed_image_digest | Not specified |
| executed_image_id | Not specified |
| system | eight DGX Spark GB10, two switched 200G RoCEv2 subnets, MTU9000 |
| node_count | 8 |
| parallelism | TP8/EP1 |
| weight_precision | MXFP4 expert storage; Marlin W4A16; optional FP8 o_proj and draft |
| kv_dtype | Not specified |
| experiment_date | 2026-09-24 |
| measured_build | a13 |
| temperature | 0 |
| thinking | False |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | Not specified |
| cache_state | Not specified |
| peak_runtime_RAM | Not specified |
| context_limit_tokens | Not specified |
| drafter_revision | Not specified |
| concurrency | 8 |
| metric_denominator | all eight category output / total wall; not mean per-stream multiplied by 8 |
a13 cold CSV extraction C8 · aggregate output rate · conditions
| model | MiMo-V2.6-Pro-RL |
|---|---|
| model_artifact | XiaomiMiMo/MiMo-V2.6-Pro-RL |
| model_revision | 73875d00b30a89ef8cc353a0b60b0e9f9561952d |
| engine | patched SGLang with Marlin W4A16 MXFP4 experts, DFlash8 adaptive verify and FA4 |
| base_engine_revision | 06008c170dae333e22caceb51f36224071f9b5b3 |
| runtime_execution_revision | Not specified |
| executed_image_digest | Not specified |
| executed_image_id | Not specified |
| system | eight DGX Spark GB10, two switched 200G RoCEv2 subnets, MTU9000 |
| node_count | 8 |
| parallelism | TP8/EP1 |
| weight_precision | MXFP4 expert storage; Marlin W4A16; optional FP8 o_proj and draft |
| kv_dtype | Not specified |
| experiment_date | 2026-09-24 |
| measured_build | a13 |
| temperature | 0 |
| thinking | False |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | three waves, 8 requests per wave |
| cache_state | unique cache salt; prefix cache flushed before each wave |
| peak_runtime_RAM | Not specified |
| context_limit_tokens | Not specified |
| drafter_revision | Not specified |
| concurrency | 8 |
| input_length_source_label | four CSV fixtures, duplicated per wave; C8 input counts not retained |
| output_tokens_cap | 4096 |
| actual_output_tokens_group | 43908 |
| group_wall_seconds | 187.18 |
| correct | 24 |
| total | 24 |
| seed | 20260922 |
| metric_denominator | all completion usage tokens / total synchronized wave wall including prefill |
a14 cold CSV extraction C8 · aggregate output rate · conditions
| model | MiMo-V2.6-Pro-RL |
|---|---|
| model_artifact | XiaomiMiMo/MiMo-V2.6-Pro-RL |
| model_revision | 73875d00b30a89ef8cc353a0b60b0e9f9561952d |
| engine | patched SGLang with Marlin W4A16 MXFP4 experts, DFlash8 adaptive verify and FA4 |
| base_engine_revision | 06008c170dae333e22caceb51f36224071f9b5b3 |
| runtime_execution_revision | Not specified |
| executed_image_digest | Not specified |
| executed_image_id | sha256:db156b3c950a0ec9c4a7d629f7574f65c1023150bcf2d0755c45d4f2c3cab661 |
| system | eight DGX Spark GB10, two switched 200G RoCEv2 subnets, MTU9000 |
| node_count | 8 |
| parallelism | TP8/EP1 |
| weight_precision | MXFP4 expert storage; Marlin W4A16; optional FP8 o_proj and draft |
| kv_dtype | Not specified |
| experiment_date | 2026-09-24 |
| measured_build | a14 |
| temperature | 0 |
| thinking | False |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | three waves, 8 requests per wave |
| cache_state | unique cache salt; prefix cache flushed before each wave |
| peak_runtime_RAM | Not specified |
| context_limit_tokens | Not specified |
| drafter_revision | Not specified |
| concurrency | 8 |
| input_length_source_label | four CSV fixtures, duplicated per wave; C8 input counts not retained |
| output_tokens_cap | 4096 |
| actual_output_tokens_group | 43900 |
| group_wall_seconds | 143.16 |
| correct | 24 |
| total | 24 |
| seed | 20260922 |
| metric_denominator | all completion usage tokens / total synchronized wave wall including prefill |
a14 cold CSV extraction C1 · aggregate output rate · conditions
| model | MiMo-V2.6-Pro-RL |
|---|---|
| model_artifact | XiaomiMiMo/MiMo-V2.6-Pro-RL |
| model_revision | 73875d00b30a89ef8cc353a0b60b0e9f9561952d |
| engine | patched SGLang with Marlin W4A16 MXFP4 experts, DFlash8 adaptive verify and FA4 |
| base_engine_revision | 06008c170dae333e22caceb51f36224071f9b5b3 |
| runtime_execution_revision | Not specified |
| executed_image_digest | Not specified |
| executed_image_id | Not specified |
| system | eight DGX Spark GB10, two switched 200G RoCEv2 subnets, MTU9000 |
| node_count | 8 |
| parallelism | TP8/EP1 |
| weight_precision | MXFP4 expert storage; Marlin W4A16; optional FP8 o_proj and draft |
| kv_dtype | Not specified |
| experiment_date | 2026-09-24 |
| measured_build | a14 |
| temperature | 0 |
| thinking | False |
| input_tokens | 1498 |
| output_tokens | Not specified |
| repetitions | four tasks x three repeats |
| cache_state | unique salt and cache flush before every request |
| peak_runtime_RAM | Not specified |
| context_limit_tokens | Not specified |
| drafter_revision | Not specified |
| concurrency | 1 |
| output_tokens_cap | 4096 |
| actual_output_tokens_group | 21952 |
| group_wall_seconds | 209.86 |
| correct | 12 |
| total | 12 |
| seed | 20260922 |
| metric_denominator | completion usage / summed client wall including prefill |
a13 GSM8K first 200 · accuracy · conditions
| model | MiMo-V2.6-Pro-RL |
|---|---|
| model_artifact | XiaomiMiMo/MiMo-V2.6-Pro-RL |
| model_revision | 73875d00b30a89ef8cc353a0b60b0e9f9561952d |
| engine | patched SGLang with Marlin W4A16 MXFP4 experts, DFlash8 adaptive verify and FA4 |
| base_engine_revision | 06008c170dae333e22caceb51f36224071f9b5b3 |
| runtime_execution_revision | Not specified |
| executed_image_digest | Not specified |
| executed_image_id | Not specified |
| system | eight DGX Spark GB10, two switched 200G RoCEv2 subnets, MTU9000 |
| node_count | 8 |
| parallelism | TP8/EP1 |
| weight_precision | MXFP4 expert storage; Marlin W4A16; optional FP8 o_proj and draft |
| kv_dtype | Not specified |
| experiment_date | 2026-09-24 |
| measured_build | a13 |
| temperature | 0 |
| thinking | False |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | Not specified |
| cache_state | Not specified |
| peak_runtime_RAM | Not specified |
| context_limit_tokens | Not specified |
| drafter_revision | Not specified |
| concurrency | Not specified |
| harness_default_concurrency | 8 |
| actual_harness_concurrency_reported | False |
| benchmark | first 200 GSM8K test questions in fixed order |
| correct | 196 |
| total | 200 |
| shots | 0 |
| output_tokens_cap | 1024 |
| answer_extraction | Answer: numeric, else last numeric; abs tolerance <1e-6 |
| wrong_ids | [93, 119, 162, 187] |
| a14_GSM8K_rerun | False |
| leaderboard_equivalent | False |
a13 5K cold prefill · prefill rate · conditions
| model | MiMo-V2.6-Pro-RL |
|---|---|
| model_artifact | XiaomiMiMo/MiMo-V2.6-Pro-RL |
| model_revision | 73875d00b30a89ef8cc353a0b60b0e9f9561952d |
| engine | patched SGLang with Marlin W4A16 MXFP4 experts, DFlash8 adaptive verify and FA4 |
| base_engine_revision | 06008c170dae333e22caceb51f36224071f9b5b3 |
| runtime_execution_revision | Not specified |
| executed_image_digest | Not specified |
| executed_image_id | Not specified |
| system | eight DGX Spark GB10, two switched 200G RoCEv2 subnets, MTU9000 |
| node_count | 8 |
| parallelism | TP8/EP1 |
| weight_precision | MXFP4 expert storage; Marlin W4A16; optional FP8 o_proj and draft |
| kv_dtype | Not specified |
| experiment_date | 2026-09-24 |
| measured_build | a13 |
| temperature | 0 |
| thinking | False |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | 3 |
| cache_state | unique cold prefix |
| peak_runtime_RAM | Not specified |
| context_limit_tokens | Not specified |
| drafter_revision | Not specified |
| concurrency | 1 |
| input_length_source_label | 5K |
| statistic | median |
| metric_denominator | TTFT based |
a13 163K cold prefill · prefill rate · conditions
| model | MiMo-V2.6-Pro-RL |
|---|---|
| model_artifact | XiaomiMiMo/MiMo-V2.6-Pro-RL |
| model_revision | 73875d00b30a89ef8cc353a0b60b0e9f9561952d |
| engine | patched SGLang with Marlin W4A16 MXFP4 experts, DFlash8 adaptive verify and FA4 |
| base_engine_revision | 06008c170dae333e22caceb51f36224071f9b5b3 |
| runtime_execution_revision | Not specified |
| executed_image_digest | Not specified |
| executed_image_id | Not specified |
| system | eight DGX Spark GB10, two switched 200G RoCEv2 subnets, MTU9000 |
| node_count | 8 |
| parallelism | TP8/EP1 |
| weight_precision | MXFP4 expert storage; Marlin W4A16; optional FP8 o_proj and draft |
| kv_dtype | Not specified |
| experiment_date | 2026-09-24 |
| measured_build | a13 |
| temperature | 0 |
| thinking | False |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | 3 |
| cache_state | unique cold prefix |
| peak_runtime_RAM | Not specified |
| context_limit_tokens | Not specified |
| drafter_revision | Not specified |
| concurrency | 1 |
| input_length_source_label | 163K |
| statistic | median |
| metric_denominator | TTFT based |
Local Spark Qwen draft; Codex source audit and rewrite. No site model benchmark.