ANTIMEME SYSTEMWORKS
ANTIMEME SYSTEMWORKSCREATIVE UTILITIES DIVISIONBENCHMARK // V1.0_BUILD

TOKS ON WINDOWS

17–22x faster on 512-byte inputs. 41–56x on long and batched inputs, where 58–78% of bytes repeat within the pass. Native Windows CPU tokenization, six-round medians against the faster measured Hugging Face API. This measures encoding, not generated tokens.

AM-RS-2026 // RUN 20261018-1834319800X3D // AVX2 // HOST SELF-REPORTEDSOURCE + EVIDENCE
01

Measured, with conditions

Short-input ratio
17–22x
512 B // ≈14% REPEATED BYTES
Token-ID parity
6,240
DISTINCT TOKENIZER/TEXT PAIRS
Timing samples
216
ACCEPTED // 9 REJECTED
Consistency audit
741
CHECKS // HARDWARE NOT ATTESTED

FIG.01Input throughput

M INPUT TOK/S
C01–C09 MAP TO TABLE CELLS // MEDIANS // AXIS STARTS AT ZERO

FIG.02Relative rate

TOKS / HF // RATIO
LONG/BATCH CELLS REUSE 58–78% OF BYTES // CONNECTING LINES ONLY GUIDE THE EYE
02

All single-worker cells

First pass // six rounds

MEDIAN [MIN–MAX] // SELECT A ROW
CellTokenizer fileInputtoksM TOK/SHF fastM TOK/SRatioxRepeated%Read as
C01zai-org/GLM-5.3-Flash512 B24.75 [24.47–25.06]1.43 [1.39–1.46]17.35≈14starting point
C02zai-org/GLM-5.3-Flash64 KiB53.53 [52.69–53.70]1.31 [1.27–1.35]40.93≈58within-pass reuse
C03zai-org/GLM-5.3-Flash32 × 4 KiB73.05 [72.07–74.98]1.58 [1.55–1.59]46.23≈78within-pass reuse
C04Qwen/Qwen3.8-27B512 B25.24 [25.23–25.79]1.17 [1.12–1.21]21.56≈14short input
C05Qwen/Qwen3.8-27B64 KiB53.53 [52.42–53.89]1.18 [1.17–1.20]45.22≈58within-pass reuse
C06Qwen/Qwen3.8-27B32 × 4 KiB73.71 [72.43–74.75]1.37 [1.37–1.39]53.77≈78within-pass reuse
C07deepseek-ai/DeepSeek-V4.1-Flash-1017512 B25.98 [25.13–27.20]1.18 [1.15–1.21]22.04≈14short input
C08deepseek-ai/DeepSeek-V4.1-Flash-101764 KiB56.10 [52.54–57.03]1.16 [1.12–1.17]48.36≈58within-pass reuse
C09deepseek-ai/DeepSeek-V4.1-Flash-101732 × 4 KiB73.70 [71.70–75.35]1.31 [1.31–1.34]56.06≈78within-pass reuse

C01 IS A CONSERVATIVE STARTING POINT, NOT A MODEL RECOMMENDATION // RATIOS COMPARE MEDIANS // RANGES ARE NOT CONFIDENCE INTERVALS

03

Method + evidence

Host
Ryzen 7 9800X3D; 8 cores / 16 logical; self-reported
Runtime
Windows 11 / CPython 3.12.14 / AVX2
Libraries
toks 0.3.2 fork-built wheel / tokenizers 0.23.2
Protocol
6 balanced orders; independent parity processes; fresh timing process per library/round
Run-start CPU
2.25–6.45%; five accepted samples
Per-trial CPU
0.39–9.96%; two accepted samples per worker; threshold 10%
Load resolution
72 / 216 first passes resolved; zero toks first passes resolved (5–13 ms)
Inputs
English prose and CPython source windows; overlaps included
Rejected attempts
7 / 9 rejected only in the repeated phase; headline sensitivity ≤1.4%
UTF-8
Non-ASCII conversion included in the first timed call for both libraries
VERIFY SAVED EVIDENCE // NO TIMING
git clone https://github.com/outerheaven199X/toks-windows-benchmark.git
cd toks-windows-benchmark
python tools/download_evidence.py --output evidence/run
$kit = 'evidence/run/toks-vs-hf-windows-v2-public'
python "$kit/audit_results.py" "$kit/runs/20261018-183431" "$kit/runs/20261018-183431/audit-local.json"

Python 3.11+ for audit commands. Expected: 741 checks, exit 0. Full reproduction instructions.

Independent review

All 576 summary values recomputed from the raw receipts with zero mismatches. A reviewer-built Linux copy matched token-ID hashes in all nine unique input cells, with ratios of 13.4–44.2x on different hardware.

The reviewer did not independently observe the Windows host or rebuild its binary. These Linux numbers are supplied reviewer evidence.

Read the independent report