Research
Sthānika AI's research studies measure how AI models perform on the tasks they are actually used for, each published with its corpus, evaluation harness and raw model responses.
We publish the corpus, the evaluation harness and the raw model responses for every study. A score you cannot reproduce is an assertion, not a measurement.
- 2026-10-19DecisionNativeBench: an evaluation set for decision-native models
29 open-weight models and hosted Jev 1.13, every one of the 5,405 typed-decision questions across 11 business domains. Jev scores 90.5%; the best open-weight model, Winnow-12B, 86.9%.
- 2026-09-24Writing India: text-to-image models flatline on Indic scripts
Nine open-weight text-to-image models across Hindi, Bengali, Marathi, Telugu and Kannada. Eight produce a flat 0% exact match on legible Indic signage.
- 2026-09-24What It Actually Said: 16 Indic speech-to-text models
Sixteen ASR models on IndicVoices across 22 languages and 357,000+ utterances. Four produce measurable but meaningless scores by translating instead of transcribing.
- 2026-09-21Bharat Knowledge Probe: what models know about India
19 models on 552 India-specific knowledge items across seven categories. Qwen3.6 27B leads at 53.2% Bharat Score.
- 2026-09-09Can a model advise a farmer? 21 models on Kisan Call Centre questions
500 real crop-advisory questions in 11 Indian languages, graded by an LLM judge on four axes.
- 2026-09-09Indic script reading gap: what vision models miss on Indian signboards
Eight vision-language models read every legible sign in 1,319 unedited photographs from across India.
- 2026-09-09The code-mixing tax: what AI models lose on romanized Hindi
16 models asked the same 1,979 questions five ways. Writing in English letters costs more than mixing languages does.
- 2026-09-09How many tokens does Telugu cost?
Thirteen tokenizers measured across five languages. Dravidian scripts cost four to five tokens per word everywhere.
- 2026-09-09MILU, re-run on the 2026 models
17 models on the complete MILU test set, 79,608 questions across 11 languages with no subsampling, with cost per rupee alongside raw accuracy.
