Yosha Core · first model

One model for India,
grounded in Indian sources.

Starting with a 10B-parameter prototype, we are building toward Yosha 100B, a foundation model designed for the complexity, scale, and diversity of India.

Indic languages built in, not translated.

First model: 10B
Flagship: 100B

Yosha 100B design

Training tokens: 2.5T
Layers: 80
Context: 256K
Vocabulary: 128K

The 10B prototype's configuration will be fixed once our AI team joins.

yosha-100b · inference

Illustrative run

Prompt · English · Education1 / 10 tokens

Solve

128K Indic-aware BPE · split shown is illustrative

Yosha 100B80 blocks · GQA 80Q / 8KV · 256K ctxtokenizing

Early blocks 0–15Middle blocks 16–63Late blocks 64–79hover a block

Answer · English—

Training plan

10B first. 100B on 2.5T tokens, when earned.

The plan: a 10B prototype proves the recipe first, once our AI team is in place. Yosha 100B follows when product usage and revenue justify the compute. It will spend two and a half trillion tokens on a hundred-billion-parameter decoder — roughly twenty-five tokens per parameter, a little past the compute-optimal point — so the base checkpoint carries Indic syntax, statute and syllabus before any industry adapter is attached.

2.5Ttraining tokensplanned budget

≈ 25×tokens per parameterpast compute-optimal

≈ 1.7passes over the corpusaveraged across the mix

Token budget

2.5T tokens

Planned targets for the Yosha 100B run, which has not started yet. The frozen configuration and training log will ship with the evaluation card.

Planned run at a glance

Model: Yosha 100Bafter the 10B prototype
Unique corpus: ≈ 1.5T tokensafter dedup and licence filter
Global batch: 16M tokens
Warmup: 20B tokensthen cosine decay
Parallelism: 3Ddata + tensor + pipeline
Checkpoint: Every 25B tokenskept for eval replay

Tokens per parameter

Compute-optimal (Chinchilla, 2022)≈ 20×

Yosha 100B≈ 25×

Training past the compute-optimal ratio costs more compute up front, and buys a stronger model at the same size, which is cheaper to serve than a bigger model of equal quality.

Training data

What the 2.5T tokens will be made of

The planned mix for the 100B run. Every shard will carry provenance and a licence tag. General English and Indic text builds broad ability; regulatory, statutory and syllabus text is upweighted well above its natural share because it is the part that has to be right.

2.5Ttraining tokens

100% accounted for · 2.5T tokens · planned mix

Industry slices will be sized to the data we license. Where a slice has less data than planned, its share shrinks and that data goes to midtraining and fine-tuning instead.

Evaluation

Benchmarks we will publish against

Held-out sets, one per industry (trade to follow), graded by industry experts as well as automatic scorers. Every score will be published next to the strongest available global and Indic models.

Our held-out sets · to be built

IndiaExam-QAEducationClass 6–12 and competitive exam question answering
IndiaBank-QABanking & FinanceRegulatory circular comprehension with citation accuracy
IndiaLawBenchLawSection identification under BNS / BNSS / BSA
SOCTriageBenchCybersecurityAlert severity agreement with analyst ground truth
IndiaClinicalBenchHealthcareGuideline-grounded answers on Indian clinical cases

Public Indic benchmarks

MILUMulti-task Indic language understanding, by AI4Bharat
IndicGenBenchGeneration across Indic languages, by Google Research

No scores yet. The held-out sets will be built with industry experts and published with the evaluation card, each result shown next to the strongest available models.