asia-fertility — results

Tokenizer cost ratios, multi-turn NIAH recall, and wall-clock latency measurements for 16 lower-resource Asian languages × 10 frontier tokenizers, with bootstrap 95% confidence intervals.

📄 Preprint (Zenodo): 10.5281/zenodo.21069313 💻 Code: Helmo21/asia-fertility 🌐 Web app: fertiscope.vercel.app

Generated by asia-fertility v0.4.0.

Configs (use with datasets)

from datasets import load_dataset

# Benchmark A — main leaderboard (160 rows: language × tokenizer × corpus)
ds   = load_dataset("Helmo21/asia-fertility", "leaderboard")

# Benchmark B — 4,000-cell NIAH multi-turn recall (v0.3)
# 16 langs × 5 models × 5 fills × 5 positions × 2 trials
niah = load_dataset("Helmo21/asia-fertility", "niah")

# Benchmark C — wall-clock latency (v0.3)
# 800 measured trials + 240 warmup, 5 models × 16 langs × ~13 calls
lat  = load_dataset("Helmo21/asia-fertility", "latency")

Files per version

File	Description
results.csv	Leaderboard — 160 rows × lang, tokenizer, fertility, premium, cost_ratio, cpt, bpt, ...
results.json / results.parquet	Same as results.csv in other formats
leaderboard.json	Nested schema-versioned view (consumed by the web app)
manifest.json	Run metadata: SHA256 of config + pinned prices + FX rates + tokenizer versions
niah_results.csv	NIAH benchmark — 4,000 rows for v0.3
latency_results.csv	Wall-clock latency benchmark — 1,040 rows (800 measured + 240 warmup) for v0.3

Columns in latency_results.csv

model, iso, script, trial, is_warmup, prompt_chars, ttft_ms, total_ms, output_chars, error

Filter is_warmup == True and rows with non-empty error before analysis.

Key findings

Cost ratios

• On cl100k_base, median cost ratio for the 11 non-Latin scripts is 8.9× (Sinhala), up to 11.7× for Burmese
• Switching from cl100k_base to o200k_base cuts Tamil from 7.61× to 1.98×, Malayalam from 8.94× to 1.95×, Burmese from 11.66× to 3.18× — no model retraining required
• Gemma-2 is the best open-weight tokenizer for South-Asian deployment (Tamil 2.58×, Hindi 1.86×, Burmese 4.80×)
• BLOOM has surprisingly low cost ratio on Indic scripts (Tamil 1.29×, Bengali 1.17×, Hindi 1.30×) but struggles with Burmese (10.03×) and Sinhala (8.27×)
• Aya Expanse underperforms despite its multilingual mandate (Tamil 6.36×, Lao 9.63×, Burmese 8.85×) — closer to cl100k than to o200k

NIAH recall (4,000-cell v0.3 grid)

• Recall collapse at 4k tokens — well inside every model's window — for gpt-4o-mini, llama-3.1-8b-instruct, qwen-2.5-7b-instruct, and deepseek-chat-v3-0324. On the 11 non-Latin Asian languages, these four models score 0–2/10 at the smallest fill level. The collapse is script-representation-driven, not context-length-driven.
• gemini-2.5-flash breaks the pattern with 84.5% pooled recall across all 16 languages × 5 fill levels (4k to 131k), demonstrating that the script-native collapse is a vendor-level training choice rather than an intrinsic limit
• "Effective context window" is language-dependent — deepseek-chat-v3-0324 recalls at 131k on Latin scripts but errors with HTTP 400 on 100% of non-Latin haystacks at the same notional size

Latency

• Pearson r between cost ratio and latency ratio = 0.314 across 75 (model, lang) cells — notably weaker than the r ≈ 0.94 reported by Somide 2026 on African languages
• Per-model breakdown: llama-3.1-8b r=+0.74, gemini-2.5-flash r=+0.52, gpt-4o-mini r=+0.43, deepseek-chat-v3-0324 r=+0.05, qwen-2.5-7b r=−0.19
• Cost is NOT a reliable proxy for UX impact in modern (2026) inference stacks. The latency-cost relationship is dominated by the provider's serving choices (continuous batching, prefill parallelism, dynamic routing) rather than by language fertility alone.

Citation

@article{pedretti2026asianlanguagetax,
  title  = {The Asian Language Tax: Quantifying the Cost, Context, and Recall Penalty of Tokenizing Lower-Resource Asian Languages in Frontier LLMs},
  author = {Pedretti, Antoine},
  year   = {2026},
  doi    = {10.5281/zenodo.21069313},
  url    = {https://doi.org/10.5281/zenodo.21069313},
}

License

CC-BY-SA 4.0 (derived from FLORES-200, Meta NLLB). Code generating these results: MIT (see Helmo21/asia-fertility).