╭────────────────────────────────────────────────────────────────────────╮
│                                                                        │
│  R I G M A R K   //   AGENT WORKLOAD RECEIPT                           │
│  BENCHMARKS LOCAL AI HOW CODING AGENTS ACTUALLY USE IT                 │
│  ▲  12/15 BASIC OUTPUT GATES PASSED — DO NOT HEADLINE                  │
│                                                                        │
├─ MODEL ────────────────────────────────────────────────────────────────┤
│                                                                        │
│  DeepSeek-V4.1-Flash-1017                                              │
│                                                                        │
├─ SINGLE STREAM ────────────────────────────────────────────────────────┤
│                                                                        │
│  WORKLOAD         tok/s       Range tok/s   Last (s)    Checks         │
│  PROSE             80.8         66.5–91.9       38.8     ✓ 5/5         │
│  CODE              99.5        98.3–111.7       73.7     ✗ 2/5         │
│  STRUCTURED*      130.2       128.1–130.8        5.9     ✓ 5/5         │
│                                                                        │
│  * Predictable JSON ceiling; not general agent performance.            │
│  Decode medians are estimates; rates include streamed reasoning.       │
│                                                                        │
├─ PREFILL ──────────────────────────────────────────────────────────────┤
│                                                                        │
│  DEPTH       Cold tok/s      Cold TTFT (s)    Replay TTFT (s)          │
│  8K               2,403               3.41               0.99          │
│  32K              2,500              13.11               1.02          │
│  64K              2,488              26.34               1.04          │
│  Medians; replay is an immediate repeat, not a proven hit.             │
│                                                                        │
├─ CAPPED CONCURRENT GENERATION ─────────────────────────────────────────┤
│                                                                        │
│  Aggregate tok/s: C1 74.2 | C2 104.7 | C4 139.7                        │
│  C4: 0/12 normal stops; 0/12 with visible output.                      │
│  Capped throughput includes reasoning; not completed agent tasks.      │
│                                                                        │
├─ APPLIANCE ────────────────────────────────────────────────────────────┤
│                                                                        │
│  Hardware: 2x NVIDIA DGX Spark (GB10, 128 GB unified memory each),     │
│    ascent-07 + ascent-08                                               │
│  Topology: TP2, one rank per Spark, RoCE over ConnectX-7               │
│  Checkpoint: bertholomus DeepSeek-V4.1-Flash-1017 EXL3 2.5 bpw @ local │
│    finger                                                              │
│  Quantisation: EXL3 2.5 bpw weights, BF16 activations                  │
│  KV cache: engine default for this lane                                │
│  Engine: TensorFold native (tensorfold-native, image                   │
│    nvcr.io/nvidia/pytorch:26.07-py3)                                   │
│                                                                        │
├─ SETTINGS ─────────────────────────────────────────────────────────────┤
│                                                                        │
│  SUITE: DEFAULT SETTINGS                                               │
│  REQUEST: {}                                                           │
│  Temperature 0.0 | top_p 1.0 | seed 20260905 | protocol 1.3.0          │
│  Decode: 5 runs, 8192-token cap | Prefill: 8K/32K/64K, 3 pairs         │
│  Concurrency: C1/C2/C4, 3 rounds, 256-token cap; workload code         │
│  Cache isolation: random-salt-per-pair                                 │
│  Comparison ID: 2026-10-two-sparks-four-models-v1                      │
│                                                                        │
├─ RECEIPT ──────────────────────────────────────────────────────────────┤
│                                                                        │
│  SOURCE     git:946394fa53db  •  clean                                 │
│  RIGMARK RELEASE: ahead of v0.0.4 (unreleased checkout)                │
│  JSON sha256:28d89818c2776f70…                                         │
│  SHARE THE CARD • LINK THE JSON RECEIPT • #RIGMARK                     │
│  github.com/alexellis/rigmark                                          │
│                                                                        │
╰────────────────────────────────────────────────────────────────────────╯