# Performance Index

This is a small, manually maintained index of useful performance expectations.
It is not the identity of the project, a universal leaderboard, or a replacement
for the linked result packets. Those packets and their JSON/log evidence are the
source of truth.

The current local verification platform is Intel Arc Pro B70. Results and
patches for other Intel Arc GPUs are welcome, but they remain clearly labeled
until matching hardware is available for independent reproduction. Testing a
contributor's patch on B70 validates the resulting B70 behavior; it does not
verify a score originally reported on different hardware.

Last manual review: **2026-09-11**.

## Read This Before Comparing Rows

- Compare results only when model revision, quantization and quality class,
  hardware/count, engine/runtime, benchmark suite and shape, cache policy,
  concurrency, and metric definition match. A larger tok/s number in a
  different row is not automatically a faster implementation.
- Most current and rapid-snapshot rows report the median generated-token rate
  for tokens 1-100 after TTFT across 12 unique cold prompts. The MiniMax rows
  instead report mean output and total throughput for a p512/n1536 benchmark.
  They must not be numerically ranked against the cold-prompt rows.
- Historical rows produced by `bench-openai-realistic-suite.py` used a
  100-event numerator over the first-to-100th timestamp span, which contains
  99 intervals. Unless a row explicitly gives both values, treat its displayed
  first-100-token score as the published legacy convention; multiply by 0.99
  for conventional interval accounting. Relative comparisons wholly within
  that historical convention are unchanged. Laguna is the first row audited
  and displayed both ways.
- A target-verified lower-precision draft does not change the declared target
  quantization, but unverified shortcuts or a lower-precision target require a
  separate row and quality label.
- Rapid snapshots are first-pass expected-performance references, not evidence
  that a model has received the same optimization depth as a dedicated lane.
- Natural run-to-run and device-to-device variance can be several percent.
  Small claims need same-window controls, repeats, or crossover evidence.
- `B70-verified` means the row was produced and quality-gated on the
  local B70 system. It does not imply upstream support, portability, warranty,
  or identical results on another host.

## B70-Verified Results

| Model / revision | Quantization | Hardware / layout | Engine / runtime | Suite / benchmark shape | Result | Quality / status | Patch, evidence, contributor |
| --- | --- | --- | --- | --- | --- | --- | --- |
| `meta-models/Muse-Glimmer-30B` family; target GGUF SHA `e63bf23b...e31770` | UD-Q8_K_XL target; pretrained BF16 DFlash draft; BF16-rounded activations into S8/F16-scale group32 WOQ | 4x Arc Pro B70 32 GB, TP4, one active generation | llama.cpp/SYCL public base `030ebb558` plus preserved private stack/final patch; fixed-N16 direct oneDNN WOQ; distributed ARGMAX/local-winner reuse | Two independent canonical prose/code/JSON 256-token packets; frozen 15-prompt cold suite, conventional timestamps 1-100 / 99 intervals; DFlash n15/p0 | **`100.088` and `100.649 tok/s`** canonical arithmetic means; pooled `100.3685`; realistic conventional first-100 median **`161.899580`**, p10 `108.573501`; full-natural median `68.585537` | **B70-verified closed/banked Q8/WOQ result**; 15/15 cache-zero; target-verified, code/JSON exact at 256, prose near-tie divergent; not BF16/lossless/universally exact; LocalMaxxing `cmss8515c00n0ms01n3begqgg` | [packet](muse-glimmer-30b-q8-woq-b70/README.md); [repro](../repro/muse-glimmer-30b-q8-woq-b70-100tps-20260813/README.md); [source](../patches/muse-glimmer-30b-b70/README.md); contributor: [Steve Seguin](https://github.com/steveseguin) |
| `poolside/Laguna-S-2.1-INT4`<br>`4bbfc285f2f8b3b6b526274c133b7b17aae6c8cb` | compressed-tensors INT4 group-32 W4A16 target; matched INT4 DFlash draft with 31 runtime E4M3FN W8A16 projections per rank; BF16 KV; target-verified speculation | 4x Arc Pro B70 32 GB, TP4+EP4, one active generation | vLLM `1a7f61fef`; XPU kernels `99886d783`; exact width-12 146/145 target graph, segmented inline DFlash 14/13 graph, decode-only 128-GRF INT4 kernel with transposed BF16 group scales, exact Q/K RMSNorm+RoPE and M12 shared-elementwise fusions | Laguna fixed realistic suite; 13 unique cold prompts, output up to 512; conventional timestamps 1-100 / 99 intervals; greedy canonical-q1 exactness; DFlash depth 11 | **125.461973 tok/s** conventional interval median; **126.729266 tok/s** historical compatibility; conventional p10 `87.142577`; full wall `56.182284` | **B70-verified, confirmed, metric-qualified**; first valid final-source suite 13/13 token-and-text exact and cache-zero, four-rank selector evidence, no warmup/benchmark retry, clean pre/post idle gates; LocalMaxxing `cms9wuuf300cqpm01t5i285tq` | [qualified packet](laguna-s-2.1-int4-b70/README.md); [record evidence](../experiments/laguna-s-2.1-xpu-b70/notes/2026-07-31-shared-elementwise-m12-record.md); [source snapshots](../patches/laguna-s-2.1-xpu-b70/README.md); contributor: [Steve Seguin](https://github.com/steveseguin) |
| `Qwen/Qwen3.8-27B-FP8`<br>`017b9c7af6b5689d5dd426a76e0bc077eb5ca20a` | official block-scaled FP8 weights unchanged; lab default-off block-W8A16 dispatch; FP16 activations/KV; target-only MTP0 | 2x ASRock Arc Pro B70 32 GB, TP2 | pinned vLLM XPU `ac7509e2b`, XPU kernels `1e90ffa672`, deterministic GDN B/A/compiler-visible state, explicit oneCCL `Work.wait()`, XPU Graph off, Inductor enabled | Complete fixed 12-prompt/six-class suite, natural 512-token cap, cache zero, raw token IDs; two fresh servers with empty compile caches; independent canaries | **34.031596 tok/s** median of `34.025180`/`34.038013`; 12/12 complete token arrays exact. Separately scoped capacity evidence: **1,112.570323 tok/s** aggregate c128 and `31.489587 tok/s` exact-32K decode. MTP rates remain withheld. | **B70-verified candidate, strict single-user qualified.** LocalMaxxing `cmtb5n45n0021qq01n13vly2h` refers to the old invalid dynamic run and withdrawal remains recommended; it is not this result. | [qualified result](../experiments/qwen38-27b-b70/notes/2026-08-28-qwen38-fp8-deterministic-eager-baseline-and-compiled-closure.md); [package](../packages/qwen38-27b-fp8-tp2-b70/README.md); [repro](../repro/qwen38-27b-fp8-vllm-tp2-asrock-b70/README.md); contributor: [Steve Seguin](https://github.com/steveseguin) |
| `RedHatAI/Qwen3.5-9B-quantized.w4a16`<br>`a398088c4228b0ae0c8c78df88fd1e4bf445f068` | publisher W4A16 (compressed-tensors INT4 group-128 weights, FP16 activations) unchanged; publisher MTP head as draft with a draft-only INT4 lm_head copy; FP16 KV; target-verified speculation | 1x Arc Pro B70 32 GB, TP1 | pinned R276 image (vLLM XPU `ac7509e2b`, XPU kernels `1e90ffa672`, row-invariant fixed-K W4A16 kernel), FULL_DECODE_ONLY XPU graphs; the scheduled-draft configuration adds two pure-Python overlays | Complete fixed 12-prompt/six-class suite, natural 512-token cap, cache zero, raw token IDs; two fresh servers; c1-c128 identity ladders; 2K-32K real-content ladder | **113.265 tok/s** class-balanced median of `113.626701`/`112.903613` at MTP depth 3 (campaign w1, 2026-09-07); `64.335 tok/s` without speculation. Separately scoped capacity evidence: **1,268.4 tok/s** aggregate c64 without speculation, 64/64 exact in both passes. One server for every batch size (draft depth scheduled by batch size, 2026-09-11): `110.7 tok/s` at one user and `1,184 tok/s` at 64 users in the same server, exact through 32 users in both runs, 18/18 exact 2K-32K. | **B70-verified, strict single-user qualified.** LocalMaxxing `cmtrhoyl1000cps01o43bhl72`. | [headline](../experiments/qwen35-9b-b70/data/2026-09-07-qwen35-9b-w4a16-matrix-result.json); [scheduled server](../experiments/qwen35-9b-b70/notes/2026-09-10-one-server-for-every-batch-size.md); [package](../packages/qwen35-9b-w4a16-b70/README.md); [repro](../repro/qwen35-9b-w4a16-b70/README.md); contributor: [Steve Seguin](https://github.com/steveseguin) |
| `0xSero/DeepSeek-V4-Flash-180B`<br>`7c360e1cd4a5168099dbc54d16d929bf6df04990`, experimental uniform-K160 artifact | FP8 block-scaled dense weights; FP4 experts; FP8 KV; unchanged K160 target with target-verified DSpark7 | 4x Arc Pro B70 32 GB, TP4+EP, one active generation | vLLM/XPU `264c7f2f7`; XPU kernels `313156737`; oneCCL `48fda4f0e`; target and draft PIECEWISE graphs | Fixed realistic suite; 12 unique cold prompts, output 128; median tokens 1-100 after TTFT; target verifier M=8 | **80.820052 tok/s** record strict-suite high; p10 `71.669556`; three-run median `78.287226`; wall full128 `67.762818` | **B70-verified closed-lane record**; 36/36 realistic rows cache-zero, 24/24 exact canaries, unchanged target verifies accepted tokens; LocalMaxxing `cmrquta9905w3lg013m5vxoqx` | [packet](deepseek-v4-flash-k160-b70/README.md); [repro](../repro/deepseek-v4-flash-k160-b70-80tps-20260718/README.md); [summary](../experiments/deepseek-v4-flash-reap-xpu-b70/data/dspark-sharded-target-argmax-record-20260718.json); contributor: [Steve Seguin](https://github.com/steveseguin) |
| `unsloth/Qwen3.6-27B-MTP-GGUF`<br>Q4_0 file identity recorded in the linked payload | GGUF Q4_0 target; Q8 target KV; native Q4_K_M DFlash draft with F16 draft KV; target-verified speculation | 1x Arc Pro B70 32 GB, one active generation | llama.cpp/SYCL base `e3546c794` plus preserved BMG-AOT Xe2 M=6 verifier, GDN snapshot-cache, and fused Q6_K draft-head/top-1 stack | Fixed realistic suite; 12 unique cold prompts, output 128; median tokens 1-100 after TTFT; native DFlash5 | **47.818818 tok/s** median; p10 `39.869534`; mean `46.638647` | **B70-verified closed-lane record**; all cached tokens zero; unchanged target verifies accepted draft tokens; matching AOT control `44.2205 tok/s`; LocalMaxxing `cmrjbx8bc02g8mj01yzz2v701` | [closure](../notes/2026-07-13-qwen27-dflash-sycl-closure.md); [evidence](../data/qwen36-27b-mtp-gguf-q4-b70-baselines/q6top1-aot-realistic128-r2-20260713.json); [record note](../experiments/qwen27-dflash-sycl-b70/notes/2026-07-13-q6k-m6-fused-top1-production-record.md); contributor: [Steve Seguin](https://github.com/steveseguin) |
| `webhie/Qwen3.6-27B-int4-AutoRound`<br>`f5750c90b3776db658594df5fe8051098226dd8e` | AutoRound INT4 W4A16 target; FP16 target compute; runtime INT8 target LM-head with BF16 scales; runtime INT4 group-128 draft LM-head with BF16 scales | 2x Arc Pro B70 32 GB, TP2, concurrency 1 | vLLM/XPU `0.20.2rc1.dev13` local stack; pinned public oneCCL parent `b52f40c` / libccl `4ceafd1`; captured draft, graph-safe FlashAttention full target graph, ReplaySSM pending/direct-output transaction fusion | Historical 12-prompt cold suite plus independent 25-prompt, six-start review; output 512; median events 1-100 / 99 intervals; ctx 2048; MTP3 | Historical **95.384868 tok/s** (`94.431019` conventional). Independent four-arm central estimate **98.766 tok/s**, arm range `98.353–101.078`; matching target-only `47.868/48.006` | **Historical row retained; current strict review failed.** 2026-08-18: deterministic ceiling `94.710 tok/s` on the 25-prompt suite (quality pass, reproduces itself); fastest non-reproducing arm `96.822`. On the 12-prompt historical suite the record was set on, those same arms score `89.766` and `94.103`, so **nothing beats the retained 95.385 like-for-like**; all-25 medians are inflated ~5 tok/s by the faster holdout prompts. Complete-token parity against a differently-configured reference is unsatisfiable at fp16. All independent arms were cache-zero and passed objective canaries, but each speculative arm differed from target-only on 25/25 prompts and same-pair fresh starts were not exact. Do not promote the fastest arm or claim robust `>100`; no new LocalMaxxing submission. | [independent validation](../experiments/qwen36-27b-autoround-int4-b70/validation-20260815/README.md); [historical repro](../repro/qwen36-27b-autoround-int4-b70/README.md); [source bundle](../patches/qwen36-27b-autoround-int4-b70/record-20260711/README.md); [historical packet](qwen36-27b-autoround-int4-b70/tp2-fp16-fullgraph-transaction-20260711.json); contributor: [Steve Seguin](https://github.com/steveseguin) |
| `ggml-org/Qwen3.6-27B-GGUF`<br>`8a7ee08e8b9bfb857107ecc25a5599d2f38b76f8` | GGUF Q8_0 target; F16 KV; no draft or speculation | 2x ASRock Arc Pro B70 32 GB, TP2, concurrency 1 | mndodd llama.cpp/SYCL `4302fb599` plus the preserved exact lab fusion/state-I/O/tail/register-direct-Q8/IMRoPE-cache/vec4-reduction/SIMD16-QK/conv-SiLU-L2 patch; oneAPI 2026.1 BMG-G31 AOT; graph/DNN/host fallback off | Qwen fixed realistic suite; raw completions; temperature 0 / seed 42; 12 unique cold prompts, output 512; median events 1-100 after TTFT | **36.347290 tok/s conventional**; `36.714434` historical compatibility; conventional p10 `35.973240`; full-512 after-TTFT `36.365074`; wall `35.903294`; TTFT `179.962 ms` | **B70-verified target-only record**; all cached tokens zero; 12/12 complete 512-token output hashes exact; recurrent conv+SiLU+Q/K-L2 fusion `+0.322%` over the prior record and `+0.494%` full-512; `+17.153%` over matched mndodd | [packet](qwen36-27b-q8-tp2-asrock-b70/README.md); [repro](../repro/qwen36-27b-q8-tp2-asrock-b70/README.md); [summary](../data/qwen36-q8-tp2-asrock-b70-20260814/summary.json); [patch](../patches/qwen36-27b-q8-tp2-asrock-b70/README.md); source: [mndodd](https://github.com/mndodd/llama.cpp/tree/intel-sycl-optimization); lab: [Steve Seguin](https://github.com/steveseguin) |
| `ggml-org/Qwen3.6-27B-GGUF`<br>`8a7ee08e8b9bfb857107ecc25a5599d2f38b76f8` | GGUF Q8_0 target; F16 KV; no draft or speculation | 2x ASRock Arc Pro B70 32 GB, TP2, concurrency 1 | mndodd llama.cpp/SYCL fork `4302fb599` plus separated low-RAM/DNN-off compile compatibility and exact-F32 two-card all-reduce patch; oneAPI 2026.1 BMG-G31 AOT | Qwen fixed realistic suite; raw completions; temperature 0 / seed 42; 12 unique cold prompts, output 128; median tokens 1-100 after TTFT | **31.025377 tok/s conventional**; `31.338765` historical 100-event compatibility; p10 conventional `30.876411`; wall full-output median `29.922846`; TTFT `179.092 ms` | **B70-verified target-only**; all cached tokens zero; 12/12 complete output hashes equal the matched upstream-derived control; fresh logits gate PPL `5.635366` vs `5.635427`, same-top `100%`; fork is `+5.836%` over that control under either accounting convention | [community packet](../community/mndodd-qwen36-27b-llamacpp-sycl/README.md); [validation](../community/mndodd-qwen36-27b-llamacpp-sycl/validation/2026-08-12-asrock-b70-validation.md); [patch](../community/mndodd-qwen36-27b-llamacpp-sycl/patches/0001-asrock-lab-lowram-dnnless-tp2.patch); source: [mndodd](https://github.com/mndodd/llama.cpp/tree/intel-sycl-optimization) |
| `unsloth/gemma-4-26B-A4B-it-GGUF`<br>HF revision not recorded in packet; exact byte-verified GGUF is identified there | UD-Q8_K_XL target/verifier; Q4_0 MTP draft; f16 KV | 1x Arc Pro B70 32 GB, one full replica, concurrency 1 | llama.cpp/SYCL `c926ad098` plus the preserved local Gemma record stack | `gemma4-26b-a4b-q8-b70-realistic-v1`; 12 unique cold prompts, output 512; median tokens 1-100 after TTFT; ctx 32768; target-verified MTP3 | **124.97714084813418 tok/s** median; p10 `103.836100`; full-output after-TTFT median `114.871070` | **B70-verified**; realistic final gate and fresh-response validity pass, all cached tokens zero, `512/512` canary rows | [packet](gemma4-26b-a4b-q8-b70/README.md); [standalone repro](../repro/gemma4-26b-a4b-q8-b70-125tps-20260701/README.md); [summary JSON](../data/gemma4-q8-gpu0-finalpostnorm-reproexact-full512-20260701T084728Z/summary.json); [patch snapshot](../patches/gemma4-26b-a4b-q8-b70/20260629-current-llamacpp-gemma-record-worktree.md); contributor: [Steve Seguin](https://github.com/steveseguin) |
| `Lasimeri/MiniMax-M2.7-int4-AutoRound`<br>model revision not recorded in packet | AutoRound INT4 W4A16, FP16 activations / FP16-family KV; no speculation | 4x Arc Pro B70 32 GB, TP4, batch 1 | vLLM/XPU `0.20.1-local`; base vLLM `c51df430`; llm-scaler `4bfc007`; XPU graph / Level Zero | Strict speed lane: p512/n1536, ctx 2048, max batched tokens 512; mean of four clean long repeats | **89.314195 output tok/s**; `119.085594` total tok/s | **B70-verified, historical strict-speed reference**; exact n64/n256 hashes, semantic suite, arithmetic repeat, and extended sixpack passed | [repro](../repro/minimax-m27-b70-89tps-20260520/README.md); [result JSON](../repro/minimax-m27-b70-89tps-20260520/results/promoted-result-20260519.json); [patch snapshots](../repro/minimax-m27-b70-89tps-20260520/patches/README.md); contributor: [Steve Seguin](https://github.com/steveseguin) |
| `Lasimeri/MiniMax-M2.7-int4-AutoRound`<br>model revision not recorded in packet | AutoRound INT4 W4A16, FP16 activations / FP16-family KV; no speculation | 4x Arc Pro B70 32 GB, TP4, one active generation | vLLM/XPU based on `c51df430`; llm-scaler `4bfc007`; XPU kernels `28e1f5e`; OpenAI-compatible endpoint | Deployable 32K endpoint; comparable strict gate p512/n1536 at ctx 2048; mean of four repeats | **83.172184 output tok/s**; `110.896246` total tok/s; warm endpoint about `83.8` output tok/s | **B70-verified, deployable reference**; strict gate passed; serves ctx 32768. Kept separate from the faster 2K strict lane | [fresh-install repro](../repro/minimax-m27-b70-110tps-ubuntu24-20260523/README.md); [summary JSON](../repro/minimax-m27-b70-110tps-ubuntu24-20260523/results/summary-20260523.json); [applied patch snapshots](../repro/minimax-m27-b70-89tps-20260520/patches/README.md); contributor: [Steve Seguin](https://github.com/steveseguin) |
| `Qwen3.6-35B-A3B` | Quark W8A8 INT8 | 4x Arc Pro B70 32 GB, TP4 | vLLM/XPU PIECEWISE forced-comm graph | strict deep gate, p512/n512 | **93.550542 output tok/s** | **B70-verified closed reference**; strict quality gate passed; LocalMaxxing `cmqq4mw4c00yfqo01gb2ucgxj` | [packet](qwen36-35b-quark-int8-b70/README.md); [evidence](../data/qwen36-ablation-prefill-safe-int8-mixed-workspace-async-deep-gate-summary-20260615a13deep2.json) |
| `unsloth/Qwen3.6-27B-MTP-GGUF` | GGUF UD-Q4_K_XL target, intrinsic MTP7, `n_min=1`, `p_min=0.65` | 1x Arc Pro B70 32 GB | llama.cpp/SYCL `fdb1db877` | fixed 12-prompt cold realistic gate, output 128 | **31.480049 tok/s** published legacy median (`31.165249` conventional) | **B70-verified model/runtime support reference**; cached tokens zero; earlier MTP3 LocalMaxxing row `30.678767`, `cmr6mn5ct0076mn01on3dnpyn` | [packet and p-min evidence](qwen36-27b-mtp-gguf-q4-b70/README.md) |
| `unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF`<br>`eea7b2be5805a5f151f8847ede8e5f9a9284bf77` | GGUF UD-Q4_K_XL; f16 KV; no speculation | 1x Arc Pro B70 32 GB | llama.cpp/SYCL server 9763 (`dec5ca557`); runner recorded clean source snapshot `fdb1db877` | `rapid-model-snapshots-b70-realistic-v1`; 12 unique cold prompts, output 128; median tokens 1-100 after TTFT; ctx 4096 | **107.483884 tok/s** median; p10 `106.897744`; wall full-output median `94.118294` | **B70-verified rapid snapshot**; realistic final gate passed, prompt cache disabled, all cached tokens zero; first-pass baseline | [packet](rapid-model-snapshots-b70/qwen3-30b-a3b-instruct-2507-udq4/README.md); [result JSON](../data/rapid-model-snapshots-b70/qwen3-30b-a3b-instruct-2507-udq4-llamacpp-faon-nocacheprompt-realistic128-20260704T193409Z.json); no result-specific source patch; contributor: [Steve Seguin](https://github.com/steveseguin) |
| `bartowski/microsoft_Phi-4-mini-instruct-GGUF`<br>`7ff82c2aaa4dde30121698a973765f39be5288c0` | GGUF Q4_K_M; f16 KV; no speculation | 1x Arc Pro B70 32 GB | llama.cpp/SYCL; runner recorded clean source snapshot `fdb1db877` | `rapid-model-snapshots-b70-realistic-v1`; 12 unique cold prompts, output 128; median tokens 1-100 after TTFT; ctx 4096 | **96.548341 tok/s** median; p10 `96.350769`; wall full-output median `91.749702` | **B70-verified rapid snapshot**; prompt caches disabled, all cached tokens zero; standalone confirmation, first-pass baseline | [packet](rapid-model-snapshots-b70/phi4-mini-instruct-gguf/README.md); [result JSON](../data/rapid-model-snapshots-b70/phi4-mini-instruct-q4km-llamacpp-faon-cacheoff-confirm-ctx4096-realistic128-20260704T224303Z.json); no result-specific source patch; contributor: [Steve Seguin](https://github.com/steveseguin) |

## Community-Contributed Intel Arc Results

| Model / quantization | Hardware / runtime | Result | Evidence boundary |
| --- | --- | ---: | --- |
| `Qwen/Qwen3.6-27B` native FP8 Safetensors | 2x Arc Pro B70, TP2; contributor Docker/vLLM recipe reproduced locally | **30.171 tok/s** median decode across 15 rows | `B70-tested`; different prompt-length benchmark, not directly comparable to fixed-suite rows; [validation](../community/dominick253-qwen36-27b-fp8-tp2-docker/STATUS.md) |
| `unsloth/Qwen3.6-27B-MTP-GGUF` Q4_K_M | 1x Arc Pro B70; llama.cpp/SYCL `15586e2d7`; intrinsic MTP2 | **38.112 tok/s** on one fixed greedy 128-token request; target-only control `25.307` | `B70-tested`; visible bytes matched, but token IDs were not retained, request `min_p` differed, and this is not a fixed-suite median; [validation](../community/dominick253-qwen36-27b-llamacpp-sycl/validation/2026-08-08-reference-lab-validation.md) |

Results from other Intel Arc configurations are welcome when they include the
exact hardware, OS, model revision, quantization, runtime identity, command,
benchmark shape, quality gate, and JSON/log evidence. A community-reported row
remains distinct from a B70-verified B70 row unless it is independently
reproduced on matching hardware.

## Portability And Other-Hardware Observations

No rows have been added yet. Portable patches and observations from other
hardware may be useful to Intel XPU work, but their original performance claims
will be labeled as contributor-reported. If such a patch is tested locally, its
B70 measurement belongs in a separate row with its own runtime identity and
evidence.

## Maintaining This Index

Update this page by hand only after reviewing the linked packet and evidence.
Do not replace a row merely because a single run is faster. Preserve distinct
rows when the model revision, quantization or quality class, runtime, hardware
count, benchmark shape, cache policy, or metric definition changes. Superseded
rows should remain discoverable in their model packet even when this compact
index advances to a newer representative result.