Sthānika AI (“of this place”) is an independent research lab in Hyderabad building small, specialist open models for meaningful problems around the world and publishing the benchmarks that measure them, with weights and datasets released openly.
Open Models
Specialist, fine-tuned open models, released with weights, datasets, benchmarks, and reports.
100% Local
Built for the place they serve and small enough to run on your own machines.
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: high-volume bounded decisions, specialist and low-resource domains, on-premise deployment.
Every project ships its weights, its curated dataset, and its expert-graded benchmark. Reproducibility is the product.
No model trains before its evaluation set exists. Every claim on this site traces to a published benchmark, powered by our open evaluation harness and benchmarks.
Sieve-27B trained in 15.68 hours on 2× A100 GPUs, Sieve-9B-v2 in 5.56 hours on 2× H100s, and Sieve-31B in 12.03 hours on 2× B200s.
// 01 — open weights
fine-tuned weights · open inference
Sieve is our family of decision models. Give one a piece of text or JSON and a question with a fixed set of options, and it returns a probability for every option — no free text, no sampling, nothing to parse. Eight open-weight models, from 0.8B to 31B.
Sieve is a family of decision models from 0.8B to 31B, built on Qwen and Gemma backbones (both Apache-2.0). The pointer-head models follow Jared Palmer's Kev (Apache-2.0), and parts of our encoding, confidence calculation, shared-prefix execution and training recipe are adapted from its code. Sieve-31B adds a trained adapter to Mark Marosi's Decider (Apache-2.0) and uses Decider's chat prompt layout unmodified. Later models add public datasets, our own generated records, and data from Decider and Denis Yarats's AutoJev (MIT).
Our own work covers training and calibrating every checkpoint, plus two scoring designs. Option-fork scores each option as its own branch, so option order cannot change the answer. Option-codes is adapted from AutoJev's option readout. We also added per-category temperatures and built one package that loads, serves and evaluates every model. Each model ships with its Decision Index results.
Scores 63.81 on the Decision Index leaderboard, tied 3rd of 174 models, against 57.62 for the same base model read the same way. Leads the family on hard reasoning and agent tool selection, at 97.8% on BFCL.
Scores 60.17 on the Decision Index, tied 5th of 149 models. The family’s most accurate model on intent detection, claim verification and RAG hallucination checks, running on a single 80 GB GPU.
The best balance of accuracy, speed and cost. Scores 50.89 and answers in 50.5 ms, twice as fast as Sieve-27B, in about 20 GB of GPU memory.
A 3.1-point calibration error, 13th best of 149 models. Only 0.5% of the answers it gives with at least 95% confidence turn out to be wrong.
2.3B parameters, reaching 92.6 macro-F1 on CLINC150 intent detection. Each option is scored on its own, so shuffling the options can never change the answer.
A LoRA adapter for unsloth/gemma-3-12b-it, fine-tuned to answer Indian farmers' crop-advisory questions in the style of Kisan Call Centre (KCC) responses — in English and 10 Indian languages. Scores 3.44 in correctness on the Indic-KCC-Agri-Advisory-Benchmark, beating the unmodified 12B base by +1.14 and edging past gemma-3-27b-it, a base model with more than 2× the parameters.
// 02 — writing
every claim traces to a published benchmark
29 open-weight models and hosted Jev 1.13, every one of the 5,405 typed-decision questions in the evaluation set. Jev scores 90.5%; the best open-weight model, Winnow-12B, 86.9%. The larger gap is not between models but between kinds of question.
Nine open-weight text-to-image models, one 500-prompt benchmark across Hindi, Bengali, Marathi, Telugu and Kannada: eight produce a flat 0% exact match on legible Indic signage. Z-Image-Turbo is the lone exception — and only on Devanagari.
Sixteen open-weight ASR models run on AI4Bharat's IndicVoices across 22 languages and 357,000+ utterances — WER measured on real audio, not self-reported paper numbers. Four models produce measurable but meaningless scores by translating, romanizing, or reformatting instead of transcribing.
19 models answered 552 BKP-500 items across seven India-specific knowledge categories under two prompt regimes and 3 repeated samples per item. Qwen3.6 27B leads at 53.2% Bharat Score; Sarvam-M 24B is the strongest India-built model.
500 real crop-advisory questions from the Kisan Call Centre, asked in 11 Indian languages and graded by an LLM judge. DeepSeek V4-Flash leads at 4.29 out of 5; among open-weight models, Gemma 3 27B and Qwen3.6 27B tie at 3.26.
Eight open-weight vision-language models read every legible sign in 1,319 unedited photographs from across India — full frames, no crops, 13 language labels. The decisive gap is not between models but between scripts: on the 914 images carrying both, every model reads the English better than the Indic text beside it.
16 AI models asked the same 1,024 general-knowledge questions and 955 maths problems five ways — English, Hinglish, Hindi, and their romanized forms. Writing in English letters costs more than mixing languages does, and the two costs compound.
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.
We benchmarked 17 models on the complete MILU test set — 79,608 questions across 11 languages, with no subsampling — and put a real cost-per-rupee lens next to raw accuracy. Qwen3.8-Max leads at 89.67%; DeepSeek V4-Flash lands 3rd at ₹2.45 per 1,000 answered questions.
// 03 — changelog
announcements · coverage · talks
// 04 — the lab
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.
more_about_the_lab →
// 05 — open door
You hold the domain knowledge and the data exhaust; we build the model. Open front door, private deep end.
$ propose_partnership→Sovereign, on-premise, open-weight deployments — evaluated in the open.
$ start_conversation→