# Occamy GGUF validation

**Benchmark scope:** These measurements use the original qwen35-tagged artifacts. The published copies have qwen2 metadata with identical tensor payloads and separately passed bounded tokenizer and load checks; see [release verification](validation/release.json). Full benchmarks were not repeated after the metadata correction.

This report covers **12 quantization types and a BF16 reference**. It uses existing measurements only. The 60-case screen per artifact contains 24 GSM8K questions, 24 HumanEval+ tasks, 6 strict JSON cases and 6 native-tool cases. **All 312 code outputs were actually executed** with the pinned base and extended tests in an isolated CPU sandbox. This is a small fixed screen, not a full official benchmark.

[Machine-readable results](validation/results.json) · [780 per-question results](validation/per_question.jsonl)

Source: [Accio-Lab/occamy-1.0 at 8f8e0e5](https://huggingface.co/Accio-Lab/occamy-1.0/tree/8f8e0e58a3c9df042be1a3fa2c191fd8047acfb8). Runtime: [llama.cpp at 972d2313](https://github.com/ggml-org/llama.cpp/tree/972d2313bc0bf0a45f634f77d95c9fb03aeab12c). Hardware: NVIDIA B200; one GPU per job; no MTP.

## Quality results

| Artifact | PPL | Δ vs direct BF16 | Math flexible | Math strict | Code base+extended | Strict JSON | Native tools | Truncated / 60 |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| BF16 | 6.238500 | 0.000000 | 18/24 | 0/24 | 23/24 | 4/6 | 6/6 | 2 |
| IQ2_M | 7.444683 | 1.206183 | 19/24 | 0/24 | 22/24 | 4/6 | 6/6 | 2 |
| Q2_K | 7.418311 | 1.179811 | 19/24 | 0/24 | 22/24 | 4/6 | 6/6 | 3 |
| IQ3_M | 6.565706 | 0.327206 | 21/24 | 0/24 | 22/24 | 4/6 | 6/6 | 0 |
| Q3_K_M | 6.740173 | 0.501673 | 21/24 | 0/24 | 22/24 | 4/6 | 6/6 | 0 |
| IQ4_XS | 6.314710 | 0.076210 | 19/24 | 0/24 | 23/24 | 4/6 | 6/6 | 1 |
| Q4_K_M | 6.242854 | 0.004354 | 19/24 | 0/24 | 24/24 | 4/6 | 6/6 | 1 |
| Q5_K_M | 6.248758 | 0.010258 | 19/24 | 0/24 | 23/24 | 4/6 | 6/6 | 1 |
| Q6_K | 6.259335 | 0.020835 | 18/24 | 0/24 | 24/24 | 4/6 | 6/6 | 1 |
| Q8_0 | 6.257434 | 0.018934 | 18/24 | 0/24 | 23/24 | 4/6 | 6/6 | 2 |
| IQ3_XS | 6.597194 | 0.358694 | 19/24 | 0/24 | 24/24 | 4/6 | 6/6 | 3 |
| IQ4_NL | 6.318163 | 0.079663 | 19/24 | 0/24 | 23/24 | 4/6 | 6/6 | 3 |
| Q4_K_S | 6.282276 | 0.043776 | 21/24 | 0/24 | 23/24 | 4/6 | 6/6 | 0 |

Math strict-match uses the unchanged required-format extractor. The prompt did not request `####`; its failures must not be read as zero mathematical ability. Flexible-extract is reported separately using the upstream extractor. JSON is parsed as a complete response; fenced JSON fails strict parsing. Truncation remains part of the denominator. Per-question JSONL contains the original assistant content and tool_calls exactly as stored, alongside outcomes, token usage, truncation, code-test status and response digests. Dataset/task IDs identify prompts; server and infrastructure metadata are excluded.

PPL uses 32 × 512-token chunks and 8,160 scored tokens from a fixed Wikitext-2 test subset. Direct BF16 PPL is 6.2385, rounded to four decimals. The paired dump contains **lossy uint16 log-probabilities**: BF16 self-replay 6.238510 versus decoded dump base 6.232933, a 0.005577 floor. Native paired deltas and KL are kept separately in JSON.

Higher PPL in low-bit variants reflects a measured quality tradeoff, not detected artifact damage. All ten newly quantized artifacts passed full dequantized finite scans of all 733 tensors, and all twelve quantized artifacts were load-tested and benchmarked. Finite/load checks do not guarantee BF16-equivalent answers.

## Engine throughput

Native warmup and five repetitions; values are mean ± standard deviation in tokens/s. Eight CPU threads, full GPU offload requested, flash attention, F16 K/V cache, batch 2048 and microbatch 512. Tokenization and sampling are excluded. End-to-end latency and first-token latency for six prompts × three repetitions are available in results.json.

| Artifact | PP512 | TG128 @512 | PP2048 | TG128 @2048 |
|---|---:|---:|---:|---:|
| BF16 | 2959.44 ± 74.64 | 192.96 ± 3.48 | 3009.48 ± 17.28 | 193.15 ± 2.79 |
| IQ2_M | 5495.64 ± 980.55 | 211.60 ± 8.84 | 5814.88 ± 229.21 | 204.74 ± 20.24 |
| Q2_K | 4806.40 ± 994.10 | 233.02 ± 6.45 | 5085.95 ± 122.89 | 227.00 ± 18.29 |
| IQ3_M | 6262.15 ± 829.11 | 206.39 ± 5.20 | 6522.41 ± 183.47 | 203.86 ± 9.48 |
| Q3_K_M | 5998.35 ± 807.62 | 232.86 ± 2.80 | 5773.11 ± 897.22 | 232.02 ± 3.48 |
| IQ4_XS | 7037.35 ± 984.59 | 227.52 ± 10.09 | 7141.30 ± 228.24 | 228.19 ± 6.50 |
| Q4_K_M | 6346.99 ± 1065.66 | 241.55 ± 3.04 | 6311.66 ± 552.28 | 238.52 ± 7.65 |
| Q5_K_M | 6388.11 ± 711.30 | 237.84 ± 9.99 | 6310.46 ± 326.16 | 238.31 ± 9.63 |
| Q6_K | 5759.14 ± 824.86 | 223.94 ± 4.21 | 5765.40 ± 418.06 | 221.35 ± 9.85 |
| Q8_0 | 6719.33 ± 1768.73 | 223.50 ± 4.42 | 7072.93 ± 340.43 | 223.21 ± 4.71 |
| IQ3_XS | 6166.29 ± 1164.82 | 203.79 ± 2.58 | 6518.86 ± 339.58 | 203.77 ± 3.59 |
| IQ4_NL | 6856.12 ± 1322.34 | 241.74 ± 12.60 | 6910.83 ± 666.88 | 244.75 ± 3.58 |
| Q4_K_S | 6305.76 ± 1593.44 | 239.67 ± 4.07 | 6584.67 ± 270.46 | 237.58 ± 6.62 |

**Comparison limits:** IQ3_XS, IQ4_NL and Q4_K_S used a shared queue across four B200 cards, with each job on one card. Different cards and simultaneous CPU/IO activity introduce confounding relative to the earlier single-card measurements. Results are screening evidence, not a controlled performance ranking.

Ten new quantizations share one BF16 input and imatrix. The reused Q4_K_M and Q8_0 artifacts have different historical conversion recipes. Differences therefore cannot be attributed to bit width alone.

**Tokenizer boundary:** Benchmark requests use qwen35, aligned to the GGUF BF16 reference. This does not mean they reproduce the HF source tokenizer. Separate small probes found the source uses NFC normalization plus qwen2 behavior and showed disagreement with qwen35. Reused Q4/Q8 receive a qwen35 runtime override without changing their artifact bytes. Existing scores are retained as measurements under that configuration; the correction and its limited validation are documented in [tokenizer review](TOKENIZER.md). Existing Q4_K_M/Q8_0 files remain unchanged; BF16 is not a publication candidate.

## Exact measured artifacts (before release metadata correction)

The SHA256 values below identify the original benchmark inputs. Published-file SHA256 values are in [SHA256SUMS-new](SHA256SUMS-new) and [release mapping](validation/release.json). Sizes are file bytes, not runtime memory. BF16 is listed as the measured reference; inclusion in this report does not imply it is uploaded with the quantized artifacts.

| Artifact filename | Bytes | SHA256 |
|---|---:|---|
| occamy-1.0-BF16.gguf | 69376636992 | bbad6c5aaf28f20a700748613f998a2c0a98b6df94766ec37bf5c4a2cbf2aadf |
| occamy-1.0-IQ2_M.gguf | 11659235328 | 93a72b1723a06b5572d6fe9bafec42dfac1ef8b93cb1f0bf7389143b085ea8d8 |
| occamy-1.0-Q2_K.gguf | 12939593728 | caed9d813887780c3b614bec677b46e81acefdf8b13a0cfd6877df4b48ca79f4 |
| occamy-1.0-IQ3_M.gguf | 15440519168 | dd0ecc62fc13004e3c1c6094559a8d45387eca8d0b843b8c3e971eca1ed479ce |
| occamy-1.0-Q3_K_M.gguf | 16764764160 | a41137554d3d58e56dc2c043650ff92dbaf12822eb8be4a7a0d05ee1e9717d0c |
| occamy-1.0-IQ4_XS.gguf | 18728777728 | c37131d9049f031f209f2190bccb482a1d75ad9f2699ef86b67d1a193dbf1eae |
| occamy-1.0-Q4_K_M.gguf | 21166757696 | ffb25f763ff9c27f5f4e2adcdef399c5654f9f840fdac33dffe7035ba8266a87 |
| occamy-1.0-Q5_K_M.gguf | 24729131008 | 43cb6af4a64e846a32b3e87502d51008a9a8123bd76f9462120a6dff7ae6c6e4 |
| occamy-1.0-Q6_K.gguf | 28514152448 | 43e1b751fe3ebc23bccb72c3183763aabba33741854264033f4e7bb56865c829 |
| occamy-1.0-Q8_0.gguf | 36903139456 | 9bd6eceae8434098c8c7ebec33e4ca0d0bd3f004c2bddf37d16cd906448fdbdd |
| occamy-1.0-IQ3_XS.gguf | 14484144128 | 126b5265fdb52dc53b69be5535c436267cd83acdc87f1820a02ade728d3f5ef4 |
| occamy-1.0-IQ4_NL.gguf | 19779278848 | 6dcf068e182a482b0bdbc493ba7154712a034825d47ebf348d9dee7bece007c0 |
| occamy-1.0-Q4_K_S.gguf | 19889903616 | d7b4dc1b3f8b3be6fa0a16eda8bfb613119afa5a7c8db49d823b59b8af988b0d |

Evaluation source pins and full metrics are in [validation/results.json](validation/results.json). No conclusion about MTP acceleration is drawn from these MTP-disabled runs.