Across languages
How does behavior change between native script, Romanized input and unfamiliar phrasing?
Language evaluationsIndian-language models, tool-using agents and reusable capability. Explore what we’ve measured, how systems are designed and the questions still open.

Three connected questions guide the work. Each needs a check that can disagree with us.
How does behavior change between native script, Romanized input and unfamiliar phrasing?
Language evaluationsWhen can a useful capability be composed, reused or withdrawn without hiding its limits?
Explore NIFCan verification reject a convincing failure, rather than simply confirm a successful process?
Agent architectureMeasured results, architecture notes and research directions—clearly distinguished, with the underlying detail one click away.
An open speech model turned into a tool that finds the language itself, and why word error rate misleads in Indian languages.
Read the workA small open model from 36 to 76 on Sarvam’s Indic OCR bench: synthetic print, real proofread pages, and long pages that still end too soon.
Read the workWhen should useful work become reusable capability? An investigation into composition, verification and reversibility.
Read the workWhen a question returns in another language, a verified earlier answer matches thinking mode at about a fifth of its tokens.
Read the workWhere a 24B model spends its time while it writes, and a measured speed-up that keeps our adapters' 4-bit base exact.
Read the workMaths, knowledge, refusal and over-refusal beside the sarvam-m base model. Read the setup, saved results and limitations together.
Read the workHow generation, task state, tools and permissions fit together. Implemented mechanisms are not a guarantee of task success.
Read the workThe distinct roles of models, retrieval, tool protocols and reusable capability. An architectural comparison, not a leaderboard.
Read the workThe historical evaluation, with language slices, a corrected baseline and regressions preserved. Superseded by the current report.
Read the workFollow the ideas, design decisions and open questions shaping Aditya. These are notes from a build, not a list of promises.
A research direction for composing, checking and reusing capability—not a claim of established general intelligence.
Read the noteThe gap between proposing a function call and completing useful engineering work.
Read the noteTask records, checks and uncertainty belong alongside the answer.
Read the noteA module can exist before its usefulness is established. A passing evaluation does not cover every real-world task.
Model, baseline, task, sample counts and evaluation settings belong beside a result.
Regressions, refusal tradeoffs and unsuccessful cases matter as much as improvements.
Separate observations from interpretations and open hypotheses. State what remains untested.