# GLM-5.3-Flash on Two Sparks: Speed Meets Quality Gates

A pinned NVFP4 target, DFlash2 draft and community image reach sustained decode and 120124-token retrieval, while strict output failures remain visible.

- Type: Performance report
- Evidence: source_reported; site reproduced: false
- Model family: GLM
- Model: nvidia/GLM-5.3-Flash-NVFP4
- Original author: Weschera
- Original project: https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark
- Pinned document: https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/README.md
- Source revision: 281dc9a982bca17e944b322746658250a670ae73
- Source commit time (not experiment date): 2026-09-10T13:54:03Z
- Checked: 2026-10-18
- Experiment date: None
- Runtime topics: vLLM
- Hardware topics: spark, nvidia
- Reported setup: {"system": "two 128 GB DGX Spark GB10, TP2 over RoCE; text only", "node_count": 2, "parallelism": "TP2", "kv_dtype": "fp8_e4m3", "context_limit_tokens": 131072, "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": null,
    "speculative_decoding": null,
    "temperature": null,
    "top_p": null
  },
  "test_profiles": [
    {
      "label": "baseline median decode",
      "evidence_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/comparison.json#L18",
      "conditions": {
        "model": "GLM-5.3-Flash",
        "model_artifact": "nvidia/GLM-5.3-Flash-NVFP4",
        "model_revision": "423acf37583782c51c142d145aef733d72943d93",
        "runtime_execution_revision": null,
        "executed_image_digest": "sha256:4def0ef644cb2e9814136dcffd5e385e21bc594f48f3b292234051904abe85a6",
        "runtime_patch_source": "050081dc41ce6edd4d3f15fa19dc3410ba4210e3",
        "engine": "vLLM community GB10 image with Marlin W4A16 MoE",
        "system": "two 128 GB DGX Spark GB10, TP2 over RoCE; text only",
        "node_count": 2,
        "parallelism": "TP2",
        "weight_precision": "NVIDIA NVFP4 storage; Marlin W4A16 MoE compute",
        "kv_dtype": "fp8_e4m3",
        "KV_allocation_GiB": 6,
        "context_limit_tokens": 131072,
        "configured_max_requests": 1,
        "concurrency": 1,
        "prefill_batch_tokens": 128,
        "enforce_eager": true,
        "cache_state": "prefix caching disabled",
        "thinking": "custom forced-OFF template; does not qualify as native OFF",
        "temperature": 0,
        "top_p": 1,
        "seed": 37,
        "repetitions": "four different sustained prompts in one boot per arm",
        "experiment_date": null,
        "metric_denominator": "completion tokens / (wall minus measured TTFT); estimated decode",
        "output_tokens": null,
        "actual_output_tokens": null,
        "peak_runtime_RAM": null,
        "draft_tokens": 0,
        "drafter": null,
        "drafter_revision": null,
        "input_tokens": [
          78,
          92
        ],
        "actual_completion_tokens_by_prompt": [
          4356,
          2355,
          10853,
          2238
        ],
        "all_speed_requests_natural_stop": true,
        "quality_passed": false
      }
    },
    {
      "label": "dflash2-7 median decode",
      "evidence_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/comparison.json#L98",
      "conditions": {
        "model": "GLM-5.3-Flash",
        "model_artifact": "nvidia/GLM-5.3-Flash-NVFP4",
        "model_revision": "423acf37583782c51c142d145aef733d72943d93",
        "runtime_execution_revision": null,
        "executed_image_digest": "sha256:4def0ef644cb2e9814136dcffd5e385e21bc594f48f3b292234051904abe85a6",
        "runtime_patch_source": "050081dc41ce6edd4d3f15fa19dc3410ba4210e3",
        "engine": "vLLM community GB10 image with Marlin W4A16 MoE",
        "system": "two 128 GB DGX Spark GB10, TP2 over RoCE; text only",
        "node_count": 2,
        "parallelism": "TP2",
        "weight_precision": "NVIDIA NVFP4 storage; Marlin W4A16 MoE compute",
        "kv_dtype": "fp8_e4m3",
        "KV_allocation_GiB": 6,
        "context_limit_tokens": 131072,
        "configured_max_requests": 1,
        "concurrency": 1,
        "prefill_batch_tokens": 128,
        "enforce_eager": true,
        "cache_state": "prefix caching disabled",
        "thinking": "custom forced-OFF template; does not qualify as native OFF",
        "temperature": 0,
        "top_p": 1,
        "seed": 37,
        "repetitions": "four different sustained prompts in one boot per arm",
        "experiment_date": null,
        "metric_denominator": "completion tokens / (wall minus measured TTFT); estimated decode",
        "output_tokens": null,
        "actual_output_tokens": null,
        "peak_runtime_RAM": null,
        "draft_tokens": 7,
        "drafter": "incoai/GLM-5.3-Flash-DFlash2",
        "drafter_revision": "bf582e4eacc1810f76656d1811693ff6c6737d2a",
        "input_tokens": [
          78,
          92
        ],
        "actual_completion_tokens_by_prompt": [
          7396,
          2272,
          2292,
          2158
        ],
        "all_speed_requests_natural_stop": true,
        "quality_passed": false
      }
    },
    {
      "label": "dflash2-3 median decode",
      "evidence_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/comparison.json#L173",
      "conditions": {
        "model": "GLM-5.3-Flash",
        "model_artifact": "nvidia/GLM-5.3-Flash-NVFP4",
        "model_revision": "423acf37583782c51c142d145aef733d72943d93",
        "runtime_execution_revision": null,
        "executed_image_digest": "sha256:4def0ef644cb2e9814136dcffd5e385e21bc594f48f3b292234051904abe85a6",
        "runtime_patch_source": "050081dc41ce6edd4d3f15fa19dc3410ba4210e3",
        "engine": "vLLM community GB10 image with Marlin W4A16 MoE",
        "system": "two 128 GB DGX Spark GB10, TP2 over RoCE; text only",
        "node_count": 2,
        "parallelism": "TP2",
        "weight_precision": "NVIDIA NVFP4 storage; Marlin W4A16 MoE compute",
        "kv_dtype": "fp8_e4m3",
        "KV_allocation_GiB": 6,
        "context_limit_tokens": 131072,
        "configured_max_requests": 1,
        "concurrency": 1,
        "prefill_batch_tokens": 128,
        "enforce_eager": true,
        "cache_state": "prefix caching disabled",
        "thinking": "custom forced-OFF template; does not qualify as native OFF",
        "temperature": 0,
        "top_p": 1,
        "seed": 37,
        "repetitions": "four different sustained prompts in one boot per arm",
        "experiment_date": null,
        "metric_denominator": "completion tokens / (wall minus measured TTFT); estimated decode",
        "output_tokens": null,
        "actual_output_tokens": null,
        "peak_runtime_RAM": null,
        "draft_tokens": 3,
        "drafter": "incoai/GLM-5.3-Flash-DFlash2",
        "drafter_revision": "bf582e4eacc1810f76656d1811693ff6c6737d2a",
        "input_tokens": [
          78,
          92
        ],
        "actual_completion_tokens_by_prompt": [
          2822,
          2194,
          6331,
          2105
        ],
        "all_speed_requests_natural_stop": true,
        "quality_passed": false
      }
    }
  ],
  "recipe_parameters": null,
  "deployment_options": []
}
```

## Reading notes

- A source document and its selected result points are not independent experiment counts.

## Limits

- No independent Run AI Here model test.
- Executed identity, timing definition, cache state, repetition unit and missingness must be read per result.
- No cross-hardware or unmatched model score ranking.

## License declaration

{"declared_spdx": null, "scope": "Attribution and finite factual interpretation only; models and dependencies retain their own terms."}


## Editorial answer

The selected seven-token draft arm is faster than the baseline in this source, but every arm retains quality_passed=false. The fastest three-token diagnostic failed the response contract before code execution. This is a text-only community experiment with useful negative results.

By Run AI Here editorial · 2026-10-18

Local Spark Qwen writing draft; Codex source audit and rewrite. No site model benchmark.

## What is actually deployed?

The target and DFlash2 draft have separate full weight revisions, and the community image is pinned by digest. NVFP4 storage uses Marlin W4A16 MoE compute on GB10. The text-only TP2/RoCE setup is eager, uses FP8 E4M3 KV with an explicit 6 GiB allocation, a 131072 window and one request slot. It is not an NVIDIA-published GB10 production recipe, and the target export has no MTP head.

Sources: https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/README.md, https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/SOURCE_MANIFEST.json

## What workload defines the speed values?

Each arm has four naturally stopping requests: two English technical tutorials and two Korean essays. Input lengths are 78 or 92 tokens and generated lengths vary greatly. Median estimated decode excludes measured TTFT. One boot per arm and different output texts prevent a matched task-time or uncertainty claim. Saved request controls include temperature zero, top-p one and seed 37.

Sources: https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/comparison.json, https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/dflash2-7/speed-prose-0-request.json

## Why keep false quality flags?

The depth-7 arm passes the executed semantic checks but still returns a fenced block where bare Python was required and explanation where exactly 703 was required. The baseline and depth-3 arm fail the early one-code-block contract; later code checks are skipped rather than proven algorithmically wrong. All quality_passed flags remain false.

Sources: https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/comparison.json, https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/dflash2-7/quality-report.json

## What do context and thinking checks establish?

All arms retrieve three beginning/middle/end markers from 120124 actual input tokens. That is smaller than the configured window and is not broad long-document reasoning. The custom forced-OFF template keeps maximum-effort system wording and leaks planning into final text, so native OFF is unqualified. The final restart has a health probe, not a repeated quality suite; whole-model peak RAM is unreported.

Sources: https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/README.md, https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/REPRODUCE.md

## Selected source-reported figures

Each point retains its own configuration, measurement unit and source. No site reproduction or cross-source ranking.

```json
[
  {
    "label": "baseline median decode",
    "metric": "decode_rate",
    "value": 15.181243819412654,
    "unit": "tokens/s",
    "statistic": "median of four different prompts in one arm boot; not repeated identical task",
    "uncertainty": null,
    "conditions": {
      "model": "GLM-5.3-Flash",
      "model_artifact": "nvidia/GLM-5.3-Flash-NVFP4",
      "model_revision": "423acf37583782c51c142d145aef733d72943d93",
      "runtime_execution_revision": null,
      "executed_image_digest": "sha256:4def0ef644cb2e9814136dcffd5e385e21bc594f48f3b292234051904abe85a6",
      "runtime_patch_source": "050081dc41ce6edd4d3f15fa19dc3410ba4210e3",
      "engine": "vLLM community GB10 image with Marlin W4A16 MoE",
      "system": "two 128 GB DGX Spark GB10, TP2 over RoCE; text only",
      "node_count": 2,
      "parallelism": "TP2",
      "weight_precision": "NVIDIA NVFP4 storage; Marlin W4A16 MoE compute",
      "kv_dtype": "fp8_e4m3",
      "KV_allocation_GiB": 6,
      "context_limit_tokens": 131072,
      "configured_max_requests": 1,
      "concurrency": 1,
      "prefill_batch_tokens": 128,
      "enforce_eager": true,
      "cache_state": "prefix caching disabled",
      "thinking": "custom forced-OFF template; does not qualify as native OFF",
      "temperature": 0,
      "top_p": 1,
      "seed": 37,
      "repetitions": "four different sustained prompts in one boot per arm",
      "experiment_date": null,
      "metric_denominator": "completion tokens / (wall minus measured TTFT); estimated decode",
      "output_tokens": null,
      "actual_output_tokens": null,
      "peak_runtime_RAM": null,
      "draft_tokens": 0,
      "drafter": null,
      "drafter_revision": null,
      "input_tokens": [
        78,
        92
      ],
      "actual_completion_tokens_by_prompt": [
        4356,
        2355,
        10853,
        2238
      ],
      "all_speed_requests_natural_stop": true,
      "quality_passed": false
    },
    "evidence_state": "source_reported",
    "site_reproduced": false,
    "source_document_id": "glm-weschera-extra3",
    "source_document_sha256": "36dc46ce255d484340558a64b6b45ce2b79f644471d91cbe6dc0e7a82de2d461",
    "source_line": 18,
    "evidence_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/comparison.json#L18"
  },
  {
    "label": "dflash2-7 median decode",
    "metric": "decode_rate",
    "value": 20.50006471628756,
    "unit": "tokens/s",
    "statistic": "median of four different prompts in one arm boot; not repeated identical task",
    "uncertainty": null,
    "conditions": {
      "model": "GLM-5.3-Flash",
      "model_artifact": "nvidia/GLM-5.3-Flash-NVFP4",
      "model_revision": "423acf37583782c51c142d145aef733d72943d93",
      "runtime_execution_revision": null,
      "executed_image_digest": "sha256:4def0ef644cb2e9814136dcffd5e385e21bc594f48f3b292234051904abe85a6",
      "runtime_patch_source": "050081dc41ce6edd4d3f15fa19dc3410ba4210e3",
      "engine": "vLLM community GB10 image with Marlin W4A16 MoE",
      "system": "two 128 GB DGX Spark GB10, TP2 over RoCE; text only",
      "node_count": 2,
      "parallelism": "TP2",
      "weight_precision": "NVIDIA NVFP4 storage; Marlin W4A16 MoE compute",
      "kv_dtype": "fp8_e4m3",
      "KV_allocation_GiB": 6,
      "context_limit_tokens": 131072,
      "configured_max_requests": 1,
      "concurrency": 1,
      "prefill_batch_tokens": 128,
      "enforce_eager": true,
      "cache_state": "prefix caching disabled",
      "thinking": "custom forced-OFF template; does not qualify as native OFF",
      "temperature": 0,
      "top_p": 1,
      "seed": 37,
      "repetitions": "four different sustained prompts in one boot per arm",
      "experiment_date": null,
      "metric_denominator": "completion tokens / (wall minus measured TTFT); estimated decode",
      "output_tokens": null,
      "actual_output_tokens": null,
      "peak_runtime_RAM": null,
      "draft_tokens": 7,
      "drafter": "incoai/GLM-5.3-Flash-DFlash2",
      "drafter_revision": "bf582e4eacc1810f76656d1811693ff6c6737d2a",
      "input_tokens": [
        78,
        92
      ],
      "actual_completion_tokens_by_prompt": [
        7396,
        2272,
        2292,
        2158
      ],
      "all_speed_requests_natural_stop": true,
      "quality_passed": false
    },
    "evidence_state": "source_reported",
    "site_reproduced": false,
    "source_document_id": "glm-weschera-extra3",
    "source_document_sha256": "36dc46ce255d484340558a64b6b45ce2b79f644471d91cbe6dc0e7a82de2d461",
    "source_line": 98,
    "evidence_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/comparison.json#L98"
  },
  {
    "label": "dflash2-3 median decode",
    "metric": "decode_rate",
    "value": 24.77753188791761,
    "unit": "tokens/s",
    "statistic": "median of four different prompts in one arm boot; not repeated identical task",
    "uncertainty": null,
    "conditions": {
      "model": "GLM-5.3-Flash",
      "model_artifact": "nvidia/GLM-5.3-Flash-NVFP4",
      "model_revision": "423acf37583782c51c142d145aef733d72943d93",
      "runtime_execution_revision": null,
      "executed_image_digest": "sha256:4def0ef644cb2e9814136dcffd5e385e21bc594f48f3b292234051904abe85a6",
      "runtime_patch_source": "050081dc41ce6edd4d3f15fa19dc3410ba4210e3",
      "engine": "vLLM community GB10 image with Marlin W4A16 MoE",
      "system": "two 128 GB DGX Spark GB10, TP2 over RoCE; text only",
      "node_count": 2,
      "parallelism": "TP2",
      "weight_precision": "NVIDIA NVFP4 storage; Marlin W4A16 MoE compute",
      "kv_dtype": "fp8_e4m3",
      "KV_allocation_GiB": 6,
      "context_limit_tokens": 131072,
      "configured_max_requests": 1,
      "concurrency": 1,
      "prefill_batch_tokens": 128,
      "enforce_eager": true,
      "cache_state": "prefix caching disabled",
      "thinking": "custom forced-OFF template; does not qualify as native OFF",
      "temperature": 0,
      "top_p": 1,
      "seed": 37,
      "repetitions": "four different sustained prompts in one boot per arm",
      "experiment_date": null,
      "metric_denominator": "completion tokens / (wall minus measured TTFT); estimated decode",
      "output_tokens": null,
      "actual_output_tokens": null,
      "peak_runtime_RAM": null,
      "draft_tokens": 3,
      "drafter": "incoai/GLM-5.3-Flash-DFlash2",
      "drafter_revision": "bf582e4eacc1810f76656d1811693ff6c6737d2a",
      "input_tokens": [
        78,
        92
      ],
      "actual_completion_tokens_by_prompt": [
        2822,
        2194,
        6331,
        2105
      ],
      "all_speed_requests_natural_stop": true,
      "quality_passed": false
    },
    "evidence_state": "source_reported",
    "site_reproduced": false,
    "source_document_id": "glm-weschera-extra3",
    "source_document_sha256": "36dc46ce255d484340558a64b6b45ce2b79f644471d91cbe6dc0e7a82de2d461",
    "source_line": 173,
    "evidence_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/comparison.json#L173"
  }
]
```

## Supplemental source identities

```json
{
  "supplemental_sources": [
    {
      "document_id": "glm-weschera-extra0",
      "parent_record_id": "glm53-flash-nvfp4-dflash-quality",
      "publisher": "Weschera",
      "repository": "Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark",
      "revision": "281dc9a982bca17e944b322746658250a670ae73",
      "document_path": "docs/RESULTS.md",
      "pinned_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/docs/RESULTS.md",
      "document_sha256": "ff3e50dd4f7a65232e37c01e4dd79ca5bee10c9ba66dbbec0f2cf1a7b5a4c3d5",
      "last_commit_at": "2026-09-10T13:54:03Z",
      "checked_on": "2026-10-18",
      "public_URL_is_immutable": true,
      "revision_kind": "Git document revision; not necessarily the executed runtime or weights"
    },
    {
      "document_id": "glm-weschera-extra1",
      "parent_record_id": "glm53-flash-nvfp4-dflash-quality",
      "publisher": "Weschera",
      "repository": "Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark",
      "revision": "281dc9a982bca17e944b322746658250a670ae73",
      "document_path": "REPRODUCE.md",
      "pinned_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/REPRODUCE.md",
      "document_sha256": "89dd3e9334b41901f2ae8dc1d92ec8437e2ed90c143e4026c28c0978242f13c2",
      "last_commit_at": "2026-09-10T13:54:03Z",
      "checked_on": "2026-10-18",
      "public_URL_is_immutable": true,
      "revision_kind": "Git document revision; not necessarily the executed runtime or weights"
    },
    {
      "document_id": "glm-weschera-extra2",
      "parent_record_id": "glm53-flash-nvfp4-dflash-quality",
      "publisher": "Weschera",
      "repository": "Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark",
      "revision": "281dc9a982bca17e944b322746658250a670ae73",
      "document_path": "SOURCE_MANIFEST.json",
      "pinned_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/SOURCE_MANIFEST.json",
      "document_sha256": "7065d7de93ef78feb6023048fefde4506c097148cfc2733bfe65257bcbac1692",
      "last_commit_at": "2026-09-10T13:54:03Z",
      "checked_on": "2026-10-18",
      "public_URL_is_immutable": true,
      "revision_kind": "Git document revision; not necessarily the executed runtime or weights"
    },
    {
      "document_id": "glm-weschera-extra3",
      "parent_record_id": "glm53-flash-nvfp4-dflash-quality",
      "publisher": "Weschera",
      "repository": "Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark",
      "revision": "281dc9a982bca17e944b322746658250a670ae73",
      "document_path": "evidence/text128-dflash-resume/comparison.json",
      "pinned_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/comparison.json",
      "document_sha256": "36dc46ce255d484340558a64b6b45ce2b79f644471d91cbe6dc0e7a82de2d461",
      "last_commit_at": "2026-09-10T13:54:03Z",
      "checked_on": "2026-10-18",
      "public_URL_is_immutable": true,
      "revision_kind": "Git document revision; not necessarily the executed runtime or weights"
    },
    {
      "document_id": "glm-weschera-extra5",
      "parent_record_id": "glm53-flash-nvfp4-dflash-quality",
      "publisher": "Weschera",
      "repository": "Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark",
      "revision": "281dc9a982bca17e944b322746658250a670ae73",
      "document_path": "evidence/text128-dflash-resume/dflash2-7/quality-report.json",
      "pinned_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/dflash2-7/quality-report.json",
      "document_sha256": "239adbba20e74bd853f04af15301d6c5b4254512b5ee9e4b79a532b64bac0957",
      "last_commit_at": "2026-09-10T13:54:03Z",
      "checked_on": "2026-10-18",
      "public_URL_is_immutable": true,
      "revision_kind": "Git document revision; not necessarily the executed runtime or weights"
    },
    {
      "document_id": "glm-method-0",
      "parent_record_id": "glm53-flash-nvfp4-dflash-quality",
      "publisher": "Weschera",
      "repository": "Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark",
      "revision": "281dc9a982bca17e944b322746658250a670ae73",
      "document_path": "evidence/text128-dflash-resume/dflash2-7/speed-prose-0-request.json",
      "pinned_url": "https://github.com/Weschera/GLM-5.3-Flash-NVIDIA-NVFP4-2x-DGX-Spark/blob/281dc9a982bca17e944b322746658250a670ae73/evidence/text128-dflash-resume/dflash2-7/speed-prose-0-request.json",
      "document_sha256": "46e917cee3ce542c4b710c92cdb0f22fa98281177d0405f94e57af3f6fd89ff6",
      "last_commit_at": null,
      "checked_on": "2026-10-18",
      "public_URL_is_immutable": true,
      "revision_kind": "Git document revision; not necessarily the executed runtime or weights"
    }
  ]
}
```

## Paired source quality comparisons

```json
{
  "quality_comparisons": [
    {
      "variant": "baseline",
      "speculative_tokens": 0,
      "quality_passed": false,
      "runtime_checks_passed": true,
      "semantic_checks_passed": false,
      "long_context_passed": true,
      "strict_failures": [
        {
          "probe": "short-quality",
          "category": "semantic",
          "passed": false,
          "error": "Expected one code block"
        }
      ]
    },
    {
      "variant": "dflash2-7",
      "speculative_tokens": 7,
      "quality_passed": false,
      "runtime_checks_passed": true,
      "semantic_checks_passed": true,
      "long_context_passed": true,
      "strict_failures": [
        {
          "probe": "code",
          "requirement": "Only Python code, no Markdown fence",
          "passed": false
        },
        {
          "probe": "thinking-on",
          "requirement": "Final content exactly 703",
          "passed": false,
          "observed": "**Reasoning:** The sum of the first n positive integers is given by the formula n(n+1)/2. For n = 37:\n\nSum = 37 × 38 / 2 = 37 × 19 = 703\n\n**Answer:** 703"
        }
      ]
    },
    {
      "variant": "dflash2-3",
      "speculative_tokens": 3,
      "quality_passed": false,
      "runtime_checks_passed": true,
      "semantic_checks_passed": false,
      "long_context_passed": true,
      "strict_failures": [
        {
          "probe": "short-quality",
          "category": "semantic",
          "passed": false,
          "error": "Expected one code block"
        }
      ]
    }
  ]
}
```

## memory observations

```json
{
  "memory_observations": [
    {
      "kind": "explicit_KV_allocation",
      "value": 6,
      "unit": "GiB",
      "whole_model_peak_RAM": null
    }
  ]
}
```

## long context cases

```json
{
  "long_context_cases": [
    {
      "classification": "source_reported_three_marker_success",
      "actual_input_tokens": 120124,
      "configured_limit_tokens": 131072,
      "arms": [
        "baseline",
        "dflash2-7",
        "dflash2-3"
      ],
      "exact_all_three_markers": true,
      "full_window_tested": false,
      "general_long_context_reasoning_proven": false
    }
  ]
}
```

## quality comparisons

```json
{
  "quality_comparisons": [
    {
      "variant": "baseline",
      "speculative_tokens": 0,
      "quality_passed": false,
      "runtime_checks_passed": true,
      "semantic_checks_passed": false,
      "long_context_passed": true,
      "strict_failures": [
        {
          "probe": "short-quality",
          "category": "semantic",
          "passed": false,
          "error": "Expected one code block"
        }
      ]
    },
    {
      "variant": "dflash2-7",
      "speculative_tokens": 7,
      "quality_passed": false,
      "runtime_checks_passed": true,
      "semantic_checks_passed": true,
      "long_context_passed": true,
      "strict_failures": [
        {
          "probe": "code",
          "requirement": "Only Python code, no Markdown fence",
          "passed": false
        },
        {
          "probe": "thinking-on",
          "requirement": "Final content exactly 703",
          "passed": false,
          "observed": "**Reasoning:** The sum of the first n positive integers is given by the formula n(n+1)/2. For n = 37:\n\nSum = 37 × 38 / 2 = 37 × 19 = 703\n\n**Answer:** 703"
        }
      ]
    },
    {
      "variant": "dflash2-3",
      "speculative_tokens": 3,
      "quality_passed": false,
      "runtime_checks_passed": true,
      "semantic_checks_passed": false,
      "long_context_passed": true,
      "strict_failures": [
        {
          "probe": "short-quality",
          "category": "semantic",
          "passed": false,
          "error": "Expected one code block"
        }
      ]
    }
  ]
}
```