repo string | name string | bytes int64 | ppl string | kld string | rms string | top1 string |
|---|---|---|---|---|---|---|
AtomicChat | AD-BF16 | 162,100,000,000 | 4.528918 | 0.0 | 0.000 | 100.000 |
AtomicChat | AD-MXFP4 | 154,549,153,536 | 4.544560 | 0.156403 | 12.686 | 87.369 |
AtomicChat | AD-IQ3_M_XL | 143,643,963,136 | 4.549034 | 0.167467 | 13.004 | 86.864 |
AtomicChat | AD-IQ3_M | 135,792,226,048 | 4.569496 | 0.179802 | 13.615 | 86.317 |
AtomicChat | AD-IQ3_S | 130,759,061,248 | 4.601554 | 0.189076 | 13.988 | 85.945 |
AtomicChat | AD-IQ3_XS | 118,192,926,464 | 4.665654 | 0.206466 | 14.461 | 85.384 |
AtomicChat | AD-IQ3_XXS | 108,126,596,864 | 4.849062 | 0.249515 | 15.947 | 83.761 |
AtomicChat | AD-IQ2_M | 103,999,401,728 | 4.882177 | 0.256739 | 16.184 | 83.560 |
AtomicChat | AD-IQ2_S_XL | 96,751,644,416 | 5.140632 | 0.318703 | 18.500 | 81.461 |
AtomicChat | AD-IQ2_S | 93,396,201,216 | 5.215199 | 0.334328 | 18.996 | 81.031 |
AtomicChat | AD-IQ2_XS | 85,141,810,944 | 5.491651 | 0.394687 | 20.817 | 79.240 |
AtomicChat | AD-IQ2_XXS | 78,531,587,840 | 5.787808 | 0.454432 | 22.378 | 77.459 |
AtomicChat | AD-IQ1_M_XL | 72,760,225,536 | 6.178594 | 0.535118 | 24.695 | 75.162 |
AtomicChat | AD-IQ1_M | 70,243,643,136 | 6.381325 | 0.564087 | 25.253 | 74.547 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-Q8_K_XL | 161,869,615,520 | 4.538072 | 0.0 | 0.000 | 100.000 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-Q4_K_XL | 155,095,241,120 | 4.528300 | 0.155725 | 12.648 | 87.470 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-IQ4_XS | 136,662,446,656 | 4.556467 | 0.177894 | 13.509 | 86.464 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-IQ4_NL | 136,662,446,656 | 4.556467 | 0.177894 | 13.509 | 86.464 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-Q3_K_XL | 128,206,729,792 | 4.651776 | 0.198115 | 14.256 | 85.646 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-Q3_K_M | 128,078,484,032 | 4.650647 | 0.196726 | 14.172 | 85.659 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-IQ3_S | 116,069,339,712 | 4.898282 | 0.256514 | 16.461 | 83.441 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-IQ3_XXS | 104,207,848,032 | 4.910441 | 0.261009 | 16.452 | 83.463 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-Q2_K_XL | 96,832,508,352 | 5.159374 | 0.321629 | 18.609 | 80.924 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-IQ2_M | 90,926,928,288 | 5.408068 | 0.370020 | 19.996 | 79.498 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-IQ2_XXS | 90,860,736,928 | 5.415687 | 0.370894 | 19.988 | 79.456 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-IQ1_M | 86,901,313,952 | 5.685262 | 0.430815 | 21.596 | 77.751 |
unsloth/DeepSeek-V4-Flash-0731-GGUF | UD-IQ1_S | 82,539,237,792 | 5.973430 | 0.486320 | 23.077 | 76.151 |
prometheusAIR/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-IQ2_XS | 96,556,502,624 | 5.122299 | 0.316248 | 18.397 | 81.585 |
prometheusAIR/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-IQ2_XXS | 89,283,036,320 | 5.325496 | 0.360672 | 19.786 | 80.194 |
prometheusAIR/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-IQ2_XXS-slim | 84,652,524,704 | 5.516461 | 0.398412 | 20.897 | 79.133 |
DevQuasar/deepseek-ai.DeepSeek-V4-Flash-0731-GGUF | MXFP4_MOE | 154,991,538,784 | 4.525035 | 0.156042 | 12.641 | 87.508 |
DevQuasar/deepseek-ai.DeepSeek-V4-Flash-0731-GGUF | Q2_K | 103,063,965,952 | 9.112669 | 0.968548 | 32.555 | 64.429 |
DevQuasar/deepseek-ai.DeepSeek-V4-Flash-0731-GGUF | Q3_K_M | 135,318,973,920 | 4.913715 | 0.268676 | 16.406 | 82.572 |
DevQuasar/deepseek-ai.DeepSeek-V4-Flash-0731-GGUF | Q4_K_M | 171,826,479,328 | 4.548146 | 0.155595 | 12.490 | 87.451 |
DevQuasar/deepseek-ai.DeepSeek-V4-Flash-0731-GGUF | Q5_K_M | 201,636,959,712 | 4.348547 | 0.120901 | 11.344 | 88.913 |
DevQuasar/deepseek-ai.DeepSeek-V4-Flash-0731-GGUF | Q6_K | 233,310,596,096 | null | null | null | null |
DevQuasar/deepseek-ai.DeepSeek-V4-Flash-0731-GGUF | Q8_0 | 302,161,278,752 | null | null | null | null |
bullerwins/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-IQ2_XS-Experts-Q8_0 | 87,899,450,784 | 5.477248 | 0.384823 | 20.514 | 79.532 |
bullerwins/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-IQ2_XS-IQ3_XXS-GateUp-MXFP4-Down-Q8_0 | 120,329,809,280 | 4.764581 | 0.221623 | 14.963 | 84.721 |
bullerwins/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-IQ3_S-Experts-Q8_0 | 126,856,146,336 | 4.821599 | 0.249083 | 15.347 | 85.167 |
bullerwins/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-IQ3_XXS-Experts-Q8_0 | 113,870,581,152 | 4.730260 | 0.218170 | 15.034 | 85.049 |
bullerwins/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-MXFP4_MOE-BF16 | 161,869,606,432 | 4.538072 | null | 0.000 | 100.000 |
bullerwins/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-MXFP4_MOE-Q8_0 | 156,378,344,992 | 4.527893 | 0.154969 | 12.603 | 87.553 |
bullerwins/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-imatrix | 470,338,784 | null | null | null | null |
antirez/deepseek-v4-gguf | DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731 | 86,720,111,488 | 5.602505 | 0.407899 | 20.909 | 78.982 |
antirez/deepseek-v4-gguf | DeepSeek-V4-Flash-Layers37-42Q4KExperts-OtherExpertLayersIQ2XXSGateUp-Q2KDown-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-fixed-0731 | 97,591,747,456 | 5.404102 | 0.368528 | 19.972 | 80.327 |
antirez/deepseek-v4-gguf | DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731 | 155,976,458,848 | 4.544353 | 0.155195 | 12.558 | 87.439 |
antirez/deepseek-v4-gguf | DeepSeek-V4-Flash-Q4KExperts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-imatrix-0731 | 164,633,502,592 | 4.389118 | 0.123713 | 11.374 | 89.052 |
ddh0/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-3.86bpw | 137,247,731,232 | 4.580976 | 0.182301 | 13.733 | 86.247 |
ddh0/DeepSeek-V4-Flash-0731-GGUF | DeepSeek-V4-Flash-0731-MXFP4 | 156,378,344,736 | 4.559858 | 0.157541 | 12.690 | 87.425 |
DeepSeek-V4-Flash-0731 — quantization measurements
Everything needed to reproduce, audit or extend the numbers published in AtomicChat/DeepSeek-V4-Flash-0731-GGUF: the reference logits, the evaluation corpus, the raw tool output for every quant we measured, and the parsed results.
Every GGUF of this model that we could find on the Hub was measured here — ours, unsloth's, bartowski's, ggml-org's, antirez's and others — on one machine, against one reference, with one command. Publishers normally report numbers from their own harness, which makes cross-vendor comparison meaningless. These files exist so that anyone can check ours instead of trusting them.
Measured using
8x5090
Files
| File | Size | What it is |
|---|---|---|
wiki-alt.txt |
1.29 MB | Evaluation corpus: wikitext-2 test split from Salesforce/wikitext, parquet rows concatenated |
ref5632.kld |
37.1 GB | Reference logits from the lossless AD-BF16 quant over that corpus at ctx 5632 |
RESULTS-0731.jsonl |
small | Parsed results, AtomicChat and unsloth ladders |
RIVALS-B1.jsonl, RIVALS-B2.jsonl, RIVALS-B3.jsonl |
small | Parsed results, other publishers, split by the machine that produced them |
logs/*.log |
few MB | Full unedited llama-perplexity output for every quant, nothing filtered |
The .kld file stores the reference model's full probability distribution at every scored
token position — roughly 258 KB per token at this vocabulary size. It is what makes the
KL-divergence numbers comparable: every quant is compared against these exact logits.
Result schema
{"repo": "bartowski/DeepSeek-V4-Flash-0731-GGUF",
"name": "MXFP4",
"bytes": 145678901234,
"ppl": "4.5446",
"kld": "0.156403",
"rms": "12.686",
"top1": "87.369"}
ppl is Mean PPL(Q) from the KL-divergence block, kld is Mean KLD, rms is
RMS Δp, top1 is Same top p — the share of positions where the quant picks the same
next token as the reference. Note that Mean PPL(Q) and the standalone Final estimate: PPL printed by the same tool are different aggregations and do not match; the logs
contain both.
Measurement setup
| Reference | AtomicChat/DeepSeek-V4-Flash-0731-GGUF → AD-BF16 (bit-exact with the official weights) |
| Corpus | Salesforce/wikitext, wikitext-2-raw-v1, test split, rows concatenated |
| Context | 5632, batch 5632, 51 chunks |
| llama.cpp | PR #24162, commit f180ae8b2, built with -DCMAKE_CUDA_ARCHITECTURES=120 |
| GPU | 8× RTX 5090 |
Hardware matters here, and it is not optional
The routed experts of this model are 96% of its weights and they are stored in MXFP4. llama.cpp has two paths for that format — unpack to BF16 and use a normal tensor-core matmul, or feed the packed 4-bit data into block-scaled instructions. The second is gated on compute capability ≥ 12.0, which covers consumer Blackwell only. H100 and H200 are 9.0, B200 is 10.0, B300 is 10.3; all take the first path despite having FP4 hardware.
Same file, same corpus, same commit, reference model:
| GPU | ctx 512 | ctx 5632 |
|---|---|---|
| RTX 5090 | 5.4312 | 4.5381 |
| H100 | 5.1554 | 4.3406 |
A 4–5% difference from the GPU alone. Reproducing these numbers requires consumer
Blackwell and a build that targets it — compiling for sm_90 on a 5090 gives the
H100 numbers, because the native kernel never lands in the binary.
Reproducing
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
git fetch origin pull/24162/head:dsv4 && git checkout dsv4
cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=120
cmake --build build -j --target llama-perplexity
hf download AtomicChat/dsv4-eval-artifacts --repo-type dataset --local-dir .
./build/bin/llama-perplexity \
-m <any-quant>-00001-of-*.gguf \
-f wiki-alt.txt --kl-divergence-base ref5632.kld --kl-divergence \
-ngl 99 -c 5632 -b 5632
To rebuild the reference from scratch instead of downloading it, run the same command
against AD-BF16 with only --kl-divergence-base and no --kl-divergence. Takes about
ten minutes and should print Final estimate: PPL = 4.5381.
Caveats
- Absolute values are not comparable to numbers published elsewhere. Other publishers use different corpora, context lengths and hardware. Compare within one table.
- 51 chunks at ctx 5632 gives roughly ±0.003 on mean KLD. Differences smaller than that are noise.
- Quants of derived models — expert-pruned, abliterated, distilled — are deliberately excluded. KL-divergence against this reference would measure the difference between models, not the cost of quantization.
License
MIT. Derived from deepseek-ai/DeepSeek-V4-Flash-0731. Produced by
Atomic Chat.
- Downloads last month
- -