Sthānika AIwork with us →

Targeted models for
the challenges of India.
Built on locally
rooted wisdom.

Open Models

Sthānika AI (“of this place”) is a research lab building specialist, fine-tuned open models — released with weights, datasets, benchmarks, and papers.

100% Local

Built for this place and small enough to run on your own machines.

Low-angle view of Qutub Minar, a historic Indian monument
·स्थानीय·স্থানীয়·स्थानिक·స్థానిక·உள்ளூர்·સ્થાનિક·مقامی·ಸ್ಥಳೀಯ·स्थानीय·স্থানীয়·स्थानिक·స్థానిక·உள்ளூர்·સ્થાનિક·مقامی·ಸ್ಥಳೀಯ·स्थानीय·স্থানীয়·स्थानिक·స్థానిక·உள்ளூர்·સ્થાનિક·مقامی·ಸ್ಥಳೀಯ·स्थानीय·স্থানীয়·स्थानिक·స్థానిక·உள்ளூர்·સ્થાનિક·مقامی·ಸ್ಥಳೀಯ·स्थानीय·স্থানীয়·स्थानिक·స్థానిక·உள்ளூர்·સ્થાનિક·مقامی·ಸ್ಥಳೀಯ·स्थानीय·স্থানীয়·स्थानिक·స్థానిక·உள்ளூர்·સ્થાનિક·مقامی·ಸ್ಥಳೀಯ·स्थानीय·স্থানীয়·स्थानिक·స్థానిక·உள்ளூர்·સ્થાનિક·مقامی·ಸ್ಥಳೀಯ·स्थानीय·স্থানীয়·स्थानिक·స్థానిక·உள்ளூர்·સ્થાનિક·مقامی·ಸ್ಥಳೀಯ·स्थानीय·স্থানীয়·स्थानिक·స్థానిక·உள்ளூர்·સ્થાનિક·مقامی·ಸ್ಥಳೀಯ·स्थानीय·স্থানীয়·स्थानिक·స్థానిక·உள்ளூர்·સ્થાનિક·مقامی·ಸ್ಥಳೀಯ

AI models designed for
the Indian language.

6
languages
12
open artifacts
1/2
paisa / question
100%
open weights
the research →the models →
sthanika — manifest
$ cat manifest.yaml
lab: Sthānika AI, Hyderabad
parent: PurpleTalk
mission: specialist > general
releases: [weights, data, benchmark, paper]
platform: DataBounty
serving: ~½ paisa / farmer question
$
01

Specialist beats general.

A small model fine-tuned on the right data outperforms frontier models on its task — at a fraction of the cost. We target where frontier models are structurally weak: code-mixed Indian languages, India-specific domains, on-premise deployment.

02

Open is the strategy.

Every project ships its weights, its curated dataset, and its expert-graded benchmark. Reproducibility is the product.

03

Evaluation first.

No model trains before its evaluation set exists. Every claim on this site traces to a published benchmark, powered by our platform, DataBounty.

// 01 — releases

Models & projects

~/models
~/models/crop-advisory[IN DEVELOPMENT — FIRST RELEASE]

Open Indic Crop-Advisory Model

in partnership with — name on announcement

A vision-language model that acts as a conversational agronomist. Telugu, Hindi, Marathi, Kannada — native script, romanized, and code-mixed. The farmer photographs a sick plant and asks; the model answers, grounded in vetted packages of practice — it cites, it doesn't guess. Expert-gated: agronomists grade every release against a published benchmark. Crops in v1: maize, rice, cotton. Serving cost: about half a paisa per question.

paper: comingweights: comingdataset: comingbenchmark: comingdemo: coming
~/models/intake-simulator[IN DEVELOPMENT]

Intake Simulator: behavioral models for training humans

Two fine-tuned models that teach by role-play. One plays a realistic complainant at a police station — built on a hidden case file with reveal rules, measured by leak rate over long conversations. The other plays the evaluator: grading the trainee's interview and drafted FIR against per-offence ingredient checklists under the new BNS codes, every flag citing the statute. Zero-live-risk path into high-stakes institutional AI.

paper: comingweights: comingscenario_toolkit: coming
~/models/address-intelligence[EXPLORATORY]

Indian Address Intelligence

India's addresses are prose: "behind Shiv Mandir, near old peepal tree, Gali no 4." A small model that parses, normalizes, and difficulty-scores unstructured Indian addresses — trained on the one perfect label source: where the delivery actually succeeded. The most sthānika problem there is.

paper: comingweights: comingdataset: coming
~/models/expressive-indic-speech[EXPLORATORY]

Expressive Indic Speech

Off-the-shelf Indic TTS speaks in newsreader register. Training simulators need a distressed complainant, an elderly farmer, an anxious patient. We're building an acted emotional-speech corpus in Telugu and Hindi and fine-tuning open expressive TTS on it — voices with age, dialect, and feeling.

corpus: comingweights: coming
~/models/lab-report-explainer[EXPLORATORY]

Lab-Report Explainer

enabled by an existing diagnostics-chain relationship

Reads real diagnostic lab reports — messy PDFs, multi-panel tables, footnotes that reverse a flag — and explains them to patients in plain Telugu and Hindi, with deterministic escalation on critical values. The open artifact: a synthetic benchmark with expert-written gold summaries; the clinical fine-tune stays behind partner walls.

benchmark: coming
~/models/clinical-simulator[EXPLORATORY]

Clinical Conversation Simulator

enabled by an existing hospital-chain relationship

The intake-simulator engine, re-skinned for medicine: a model that plays the patient — withholding the key symptom until the right question is asked — for nursing and medical students to practice history-taking on. Same hidden case files, same leak-rate metric, zero real patients.

paper: comingscenario_toolkit: coming
~/models/medicine-explainer[EXPLORATORY]

Medicine Explainer + Open Indian Drug Dataset

enabled by an existing pharma relationship

Plain-vernacular counseling for any Indian medication — built on an open artifact worth more than the model: a brand↔salt↔generic mapping across India's hundred-thousand-plus drug brand names, expert-reviewed.

dataset: comingweights: coming
~/datasets_and_benchmarks
~/benchmarks/indic-agri[IN DEVELOPMENT]

The Indic Agri Benchmark

1,500+ farmer questions across four languages and three crops, with expert-graded gold answers. The evaluation set behind our crop-advisory model — released so anyone can measure any model against it, including ours.

dataset: comingpaper: coming
~/benchmarks/tokenizer-fertility[PUBLISHED]

The Tokenizer Fertility Suite

The sentence sets and scripts behind our tokenizer study — a permanent regression benchmark for how much Indic text costs on any tokenizer. Extend it; don't trust vendor claims.

dataset: releasedcode: releasedpaper: released
~/benchmarks/bharat-knowledge-probe[IN DEVELOPMENT]

The Bharat Knowledge Probe

Five hundred questions about the India everyone here knows: lakh–crore arithmetic, quintals and bighas, kharif and rabi, scheme names, April–March fiscal years. We score every major open model and publish the league table. Does your model know where it is?

dataset: comingleague_table: coming
~/datasets/kcc-cleaned[IN DEVELOPMENT]

KCC-Cleaned: the Kisan Call Centre corpus, usable

India's public Kisan Call Centre dataset holds millions of real farmer queries — and is famously messy. We're cleaning a slice for maize, rice, and cotton across four states, documenting every quality issue, releasing the data with the cleaning code.

dataset: comingcode: coming
~/benchmarks/india-in-the-wild[EXPLORATORY]

India-in-the-Wild: a vision benchmark

Can vision-language models read India? Hand-painted shop signage, mandi price boards, printed lab reports, handwritten Devanagari forms. A small, hard benchmark of what everyday Indian documents and streets actually look like.

dataset: comingpaper: coming
~/platform
~/platform/databounty[LIVE, INTERNAL]

The DataBounty Platform

Our data curation and evaluation platform: expert annotation workflows, blind-scoring harnesses, benchmark management, regression tracking. Experts annotate, the benchmark freezes before training, checkpoints are blind-scored. The reason our numbers can be trusted is that the pipeline producing them is a product, not a script.

$ about --platform →

// 02 — writing

Research

PUBLISHEDtokenization · reproducible

How many tokens does Telugu cost? Measuring tokenizer fertility across open models.

Telugu, Hindi, Kannada, Marathi, Hinglish, and English across Gemma, Qwen, Mistral-Tekken, and OpenAI's cl100k/o200k. Modern tokenizers sit within ~10% of each other on Indic — but the previous generation pays 2–4×. Dravidian languages cost ~4–5 tokens per word everywhere; plan serving budgets accordingly. Sentence set and scripts released.

read_postsentence_setscripts
fig 1 — tokens, same Telugu sentence
వర్షాలు ఆలస్యంగా వచ్చాయి...
Tekken
19
Gemma
20
Qwen
21
o200k
35
cl100k
86
# Dravidian scripts: ~4–5 tokens/word on every tokenizer.
# cl100k pays 4.3× Gemma.
PUBLISHEDserving economics · measured

What one H100 actually serves: measured economics of small-model deployment.

Prefill vs decode, KV-cache arithmetic, and why batching — not multiple instances — is how one GPU serves hundreds of users. The measured numbers behind the claim that a farmer's question costs half a paisa.

read_post
PREPRINT COMINGevaluation · role-play fidelity

Leak rate: evaluating whether role-play realism can be trained.

A realistic training-simulator persona must be everything an aligned model is trained not to be: withholding, inconsistent on the surface, faithful to a hidden truth. We introduce leak rate — how often a persona reveals gated facts across 30+ turn interrogations — comparing prompted frontier models against small fine-tuned ones.

abstract
~/pipeline
[01]The code-mixing tax — the same question in native script, romanized, and code-mixed: how much answer quality do models lose?in_progress
[02]Does quantization hurt Indic first? Per-language degradation at FP8 and INT4.in_progress
[03]The Indic serving-cost index — rupees per 100 answered Telugu questions, per model, refreshed quarterly.in_progress
[04]LLM-as-judge vs native speakers — can frontier models grade Telugu and Kannada answers reliably?in_progress
[05]Cross-script retrieval — does RAG survive romanization? (zyada, jyada, or jada?)in_progress
[06]Needle in a Telugu haystack — long-context recall when every word costs 4–5 tokens.in_progress
[07]The refusal gap — do safety behaviors hold across Indian languages? (Responsible-disclosure methodology.)in_progress

// 03 — changelog

News

2026-07[ANNOUNCEMENT]
How many tokens does Telugu cost? — study and benchmark suite released
Tokenizer fertility measured across six languages and five tokenizers; full sentence set and scripts published.
2026-06[ANNOUNCEMENT]
Introducing Sthānika AI — a research lab for small, specialist models
PurpleTalk launches an independent lab in Hyderabad building open, fine-tuned models for Indian languages and domains.

// 04 — the lab

About

Sthānika AI is the research lab of PurpleTalk — a twenty-year digital innovation company headquartered in Hyderabad. Where PurpleTalk builds products, Sthānika publishes research: small, specialist models built with partners who hold deep domain knowledge — seed companies, hospitals, diagnostics chains, public institutions — under one covenant: the general capability is released open; the partner's proprietary edge stays theirs.

The aim: to be the lab India's institutions trust with the models that run closest to their people.

~/purpletalk_family
Mina AI
Ello AI
xCUBE LABS
Upshot.ai
NukkadShops
YesGnome

# research ships into real products through the family

~/partners
partner logo strip — announcing soon
~/team
headshot
lab_lead
headshot
research/language
headshot
research/evaluation
headshot
eng/training
headshot
eng/serving
headshot
data/annotation

// 05 — open door

Work with us

Partners

You hold the domain knowledge and the data exhaust; we build the model. Open front door, private deep end.

$ propose_partnership→

Researchers & engineers

Small team, real GPUs, everything you build gets published.

$ see_open_roles→

Institutions & government

Sovereign, on-premise, open-weight deployments — evaluated in the open.

$ start_conversation→

// one inbox for all three

hello@sthanika.ai_