Datasets:
arm stringclasses 5
values | precision stringclasses 5
values | role stringclasses 5
values | runtime_weight_memory_gib float64 4.86 15.3 | runtime_weight_memory_bytes int64 5.22B 16.4B | memory_reduction_vs_bf16_pct float64 60.9 68.1 β | effective_bpw float64 5.08 6.25 β | tier_retention float64 0.75 0.93 β | korean_retention float64 0.64 0.9 β | accuracy_delta_vs_bf16_pp float64 -10.86 -0.31 β | ttft_ms_p50 float64 15.3 42.3 | ttft_ms_p95 float64 16.2 44 | tpot_ms_median float64 2.58 3.92 | tokens_per_s_batch1 float64 255 387 | tpot_ratio_vs_bf16 float64 0.66 1 | kernels stringclasses 5
values | kernel_mismatches float64 0 0 β | product_gate stringclasses 3
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
BF16 | BF16 weights and activations (reference) | reference | 15.268 | 16,394,145,792 | null | null | null | null | null | 15.251 | 20.28 | 3.9154 | 255.4 | 1 | BF16 dense=144 | null | reference |
A81 | NVFP4A16 body (FP4 E2M1 weights, group 16, 16-bit activations) + 46 INT8 weight-only modules (sym, group 128) + INT8 lm_head | candidate (track A) | 5.917 | 6,353,372,672 | 61.2 | 6.1732 | 0.9311 | 0.9036 | -0.31 | 15.39 | 16.241 | 3.6759 | 272 | 0.9388 | MarlinLinearKernel=46;MarlinNvFp4LinearKernel=98 | 0 | PASS |
NVFP4 W4A4 | NVFP4 weights and NVFP4 activations (W4A4) | fixed baseline for A81 | 5.977 | 6,417,801,216 | 60.9 | 6.2481 | 0.8867 | 0.8208 | -1.73 | 42.26 | 43.998 | 2.5837 | 387 | 0.6599 | FlashInferCuteDslNvFp4LinearKernel=144 | 0 | FAIL |
B81 | GPTQ INT3 body + 44 INT4 modules (group 128) | candidate (track B) | 5.153 | 5,533,509,632 | 66.2 | 5.3453 | 0.8103 | 0.6996 | -6.73 | 16.823 | 17.536 | 3.3792 | 295.9 | 0.8631 | HummingLinearKernel=100;MarlinLinearKernel=44 | 0 | FAIL |
GPTQ W3 | GPTQ INT3 weight-only, group 128 | fixed baseline for B81 | 4.864 | 5,222,719,488 | 68.1 | 5.0821 | 0.7477 | 0.6428 | -10.86 | 16.276 | 16.868 | 3.3895 | 295 | 0.8657 | HummingLinearKernel=144 | 0 | FAIL |
B200 Low-Bit Quantization Benchmarks (Qwen3-8B)
Aggregate results, recipes and serving measurements from the paper Beyond Bitwidth: Deployment-Constrained Mixed-Precision Quantization for Multilingual LLM Inference on NVIDIA Blackwell (submitted to arXiv; the link will be added once it is public).
This repository holds results only: the record of the D81 experiment (aggregate accuracy, paired differences, kernel checks, serving measurements and the A81 precision allocation). It is not the 5,168-item FINAL-81 benchmark or any other evaluation set, and it cannot be used to rerun the evaluation on its own. It contains no model weights, no evaluation items, no per-item predictions and no calibration text. The quantized checkpoints evaluated in the paper were not retained.
A separate build of the A81 recipe is released as weights in ThakiCloud/Qwen3-8B-NVFP4A16-INT8-Mixed-B200. That checkpoint was evaluated on its own, on an independent set (REL-81), and its card reports those results. They are not part of the paper's numbers and are not included in the files here.
What was tested
Base model: Qwen/Qwen3-8B (Apache-2.0). Hardware: one NVIDIA B200. Serving: vLLM 0.29.0, batch size 1.
Every quantized configuration had to pass, on a sealed evaluation set it was never tuned on:
- runtime weight memory below 5.98 GiB (measured weight-load bytes / 2^30),
- Korean retention and tier retention of at least 0.90 relative to BF16,
- overall, numeric and reasoning accuracy no more than 3 points below BF16, and format-error rate no more than 2 points above BF16,
- time per output token at most 1.1 x BF16,
- every quantized linear layer served by its intended packed kernel (no float fallback).
The confirmatory set (FINAL-81) has 5168 items: 1565 Korean, 1043 English, 1320 numeric,
870 reasoning, 300 Korean math and 70 Korean in-domain. Its SHA-256 prefix is
26a1fadc91b6. Items are not redistributed.
Main result
| Arm | Memory GiB | Tier ret. | Korean ret. | TTFT p50 ms | TPOT ms | tok/s (b1) | TPOT / BF16 | Gate |
|---|---|---|---|---|---|---|---|---|
| BF16 | 15.268 | - | - | 15.251 | 3.9154 | 255.4 | 1.0 | reference |
| A81 | 5.917 | 0.9311 | 0.9036 | 15.39 | 3.6759 | 272.0 | 0.9388 | PASS |
| NVFP4 W4A4 | 5.977 | 0.8867 | 0.8208 | 42.26 | 2.5837 | 387.0 | 0.6599 | FAIL |
| B81 | 5.153 | 0.8103 | 0.6996 | 16.823 | 3.3792 | 295.9 | 0.8631 | FAIL |
| GPTQ W3 | 4.864 | 0.7477 | 0.6428 | 16.276 | 3.3895 | 295.0 | 0.8657 | FAIL |
A81 (NVFP4A16 body, 46 INT8 modules, INT8 lm_head) was the only configuration that passed every gate:
5.917 GiB (61.2% below BF16), tier retention 0.931, Korean retention 0.904.
All 144 quantized linear layers ran on the intended kernels (98 MarlinNvFp4, 46 Marlin).
Paired comparisons
Differences are accuracy points on the same items; intervals are simultaneous at the listed z.
| Comparison | Set | Diff (pp) | CI | Status |
|---|---|---|---|---|
| A81 - NVFP4 W4A4 | FINAL-81 | +1.4 | [+0.2, +2.6] | confirmatory (preregistered) |
| B81 - GPTQ W3 | FINAL-81 | +4.1 | [+2.3, +5.9] | confirmatory (preregistered) |
| A81 - NVFP4 W4A4, Korean axis only | FINAL-81 | +1.1 | [-1.1, +3.3] | report-only |
| A81 - NVFP4 W4A4, reasoning axis only | FINAL-81 | +3.8 | [+0.3, +7.2] | report-only |
| A81 - NVFP4 W4A4 | SEL-81 | +3.4 | [+2.3, +4.6] | selection set |
| NVFP4A16 (BF16 head) - NVFP4 W4A4 | DEV-80 | +2.8 | [+1.6, +4.0] | exploratory |
| NVFP4A16 + INT8 units + INT8 head (D80 17-unit allocation) - NVFP4A16 (BF16 head) | DEV-80 | -0.3 | [-1.3, +0.8] | exploratory |
| B81 with 50% Korean calibration - B81, Korean axis | SEL-81 | +2.2 | [-0.9, +5.4] | preregistered calibration test |
| B81 with 50% Korean calibration - B81, reasoning axis | SEL-81 | -12.3 | [-16.8, -7.9] | preregistered calibration test |
How to read these numbers
- Narrow Korean margin. A81's Korean retention cleared the 0.90 gate by 0.0036 on 1565 Korean items. A different rebuild could land on the other side of the gate.
- Attribution is exploratory. On the development set, moving from NVFP4 W4A4 to 16-bit activations (NVFP4A16) accounts for much of the gain (+2.8 points [+1.6, +4.0]). Adding INT8 modules on top changed accuracy by -0.3 points [-1.3, +0.8], using an earlier 17-unit allocation. The effect of A81's own 46-unit allocation was not identified on an independent set.
- Speed trade-off. NVFP4 W4A4 had the highest batch-1 decode throughput (387 tok/s) but a TTFT of 42.3 ms and failed the quality gates. A81 decodes at 272 tok/s with a TTFT of 15.4 ms. These are batch-1 numbers; throughput at higher batch sizes was not measured.
- Not bitwise reproducible. The SEL-81 and FINAL-81 builds of A81 had different checkpoint hashes
(
9c4bce0c5e91vs7e6fcc531386). Quality variance across rebuilds was not measured. - Korean-heavy calibration did not transfer. Recalibrating B81 with 50% Korean text changed Korean accuracy by +2.2 [-0.9, +5.4] (not significant) and reasoning by -12.3 [-16.8, -7.9] on SEL-81. A81 used the default calibration mix.
Files
data/final81_arms.csvβ memory, quality retention, serving latency and gate verdict per arm.data/final81_axes.csvβ per-axis aggregate accuracy (BF16 vs quantized) per arm.data/paired_comparisons.csvβ paired differences with confidence intervals.data/kernel_verification.jsonβ kernel assignment per arm, mismatches, float fallbacks.data/run_metadata.jsonβ gates, set identity, hardware, serving configuration, build hashes.recipes/a81_allocation.jsonβ A81 module-level precision allocation (which projections are INT8).REPRODUCE.mdβ how to rebuild A81 and re-measure it.MANIFEST.sha256β file hashes.
Data sources and license
What this repository contains
Every file here is a measurement or a setting produced by ThakiCloud: per-arm and per-axis accuracy aggregates, paired differences with intervals, kernel assignments, serving latency, gate thresholds, set hashes and the A81 allocation. These files are released under CC BY 4.0. They contain no question, passage, answer option, gold label, generated output, per-item score or calibration text from any source dataset. A re-audit on 2026-11-17 scanned all 9 published files (392 string values) and found no item text and no item-keyed field. No AI-Hub data was used.
Sources of the evaluation sets (not redistributed)
FINAL-81 (5,168 items), SEL-81 (5,052) and DIAG-81 (1,998) were built by a script that pins each source to a Hub revision. The license column is what the dataset card declares at that revision.
| Source | Hugging Face repo | Revision | License (Hub card) | D81 axis | Citation |
|---|---|---|---|---|---|
| KMMLU | HAERAE-HUB/KMMLU |
d61b3f19e552c576bf5960dd24289763edc36a88 |
CC BY-ND 4.0 | Korean | Son et al., NAACL 2025 (arXiv:2402.11548) |
| KoBEST (BoolQ, COPA, HellaSwag, SentiNeg; test) | skt/kobest_v1 |
a5ea15e3ac77ed694b79f6204eb31889a2ba989f |
CC BY-SA 4.0 | Korean in-domain (report-only) | Jang et al., COLING 2022 (arXiv:2204.04541) |
| MMLU (test) | cais/mmlu |
c30699e8356da336a370243923dbaf21066bb9fe |
MIT | English | Hendrycks et al., ICLR 2021 (arXiv:2009.03300) |
| ARC (Challenge, Easy; test) | allenai/ai2_arc |
210d026faf9955653af8916fad021475a3f00453 |
CC BY-SA 4.0 | English | Clark et al., 2018 (arXiv:1803.05457) |
| HellaSwag (validation) | Rowan/hellaswag |
218ec52e09a7e7462a5400043bb9a69a41d06b76 |
none declared (see notes) | English | Zellers et al., ACL 2019 (arXiv:1905.07830) |
| BIG-Bench Hard | lukaemon/bbh |
982bb89fd79532a8ac676a61fc42eb1aeec63f99 |
none declared (see notes) | Reasoning | Suzgun et al., Findings of ACL 2023 (arXiv:2210.09261) |
| BIG-bench (41 tasks) | tasksource/bigbench |
b8ee34693560a13a9feac6aa5540449e18b7eddb |
Apache-2.0 | Reasoning | Srivastava et al., TMLR 2023 (arXiv:2206.04615) |
| MuSR | TAUR-Lab/MuSR |
7c365b439a222150f317764d4f16ae6c96d7d94a |
CC BY 4.0 | Reasoning | Sprague et al., ICLR 2024 (arXiv:2310.16049) |
HRM8K (gsm8k_test, ksm_test) |
HAERAE-HUB/HRM8K |
c360cabf8d733a82455565358b3dc965aab9ba8d |
MIT | Korean math (report-only) | Ko et al., 2025 (arXiv:2501.02448) |
| Numeric generator | ThakiCloud code, seeds 8101, 8102, 8103 | β | no third-party text | Numeric | β |
Redistribution conditions
This repository publishes none of the material above, so its CC BY 4.0 license covers only our own measurements.
If you rebuild the sets with REPRODUCE.md, download each source yourself and follow its license:
- KMMLU (CC BY-ND 4.0). You may share the dataset unchanged with attribution, but you may not share adapted material. Reformatted prompt files, subsets in our item schema and per-item outputs can count as adapted material. Do not redistribute rebuilt KMMLU files.
- KoBEST and ARC (CC BY-SA 4.0). Adapted material must credit the source and be shared under CC BY-SA 4.0. Rebuilt KoBEST or ARC items cannot be relicensed under CC BY 4.0 together with the files here.
- MuSR (CC BY 4.0), MMLU and HRM8K (MIT), BIG-bench (Apache-2.0). Attribution and the license notice are required when redistributing.
This is a summary of the declared licenses, not legal advice.
Notes from the license check (2026-11-17)
- HellaSwag. The Hub mirror declares no license at the pinned revision. The original GitHub repository
(
rowanz/hellaswag) was blocked when checked, following a DMCA notice filed by wikiHow on 2026-09-14 (recorded ingithub/dmca). HellaSwag contexts include text derived from wikiHow articles. We treat redistribution of HellaSwag items as unresolved; this repository contains none. - BIG-Bench Hard. The Hub mirror declares no license, and no license file was detected in the upstream repository
(
suzgun/BIG-Bench-Hard). Its tasks come from BIG-bench (Apache-2.0); we did not establish a separate license for BBH. - DEV-80. Two rows of the paired-comparison table use DEV-80, a set from the earlier D80 stage built by an earlier version of the same script. Its source list was not re-audited for this README.
Qwen3-8B is distributed by its authors under Apache-2.0.
Citation
@misc{han2026beyondbitwidth,
title = {Beyond Bitwidth: Deployment-Constrained Mixed-Precision Quantization for Multilingual LLM Inference on NVIDIA Blackwell},
author = {Han, Hyojung},
year = {2026},
note = {Preprint in preparation}
}
- Downloads last month
- 71