Agastya AGI v01 — F3 R2 8B Source-Token Candidate Evidence

Overview

Agastya AGI v01 is an India-developed native foundation-model research and development program undertaken by Rabbit Industries Pvt Ltd.

This record constitutes Evidence Release R2 — F3 8B Source-Token Candidate, documenting the completion of the Agastya F3 R2 candidate-training milestone on 19 October 2026.

The publication is a versioned continuation of the previously released Agastya AGI v01 Native Brain Evidence R1 archive. It provides additional company-generated technical evidence, training progress records, cryptographic identifiers, and verification metadata while preserving the historical R1 publication.

1. Training Milestone: 8 Billion Source Tokens

The Agastya F3 R2 training status record reports that the candidate-training process reached its designated target of 8,000,000,000 source training tokens, representing 100% of the specified source-token target.

The recorded completion timestamp is 19 October 2026 at 14:40:36 UTC.

The terminal training record includes:

• Source training tokens recorded: 8,000,000,000

• Target source training tokens: 8,000,000,000

• Recorded milestone completion: 100%

• Final recorded optimizer step: 7,639

• Training exposure ledger records: 10,368,986

• Terminal source-token counter reconciliation: Completed

• Final checkpoint size: 21,442,876,013 bytes

Important technical distinction: The term “8B” in this publication refers exclusively to 8 billion source tokens processed during training. It does not indicate an 8-billion-parameter model.

2. Checkpoint Integrity and Cryptographic Evidence

The supporting evidence records identify the F3 R2 final candidate checkpoint and associated training-status metadata.

The final checkpoint SHA-256 digest recorded in the checkpoint metadata is:

3e0ec187db627012e9e5eba9f73285b76f2185562b7c83841bc54e496e6325af

The training-status evidence file SHA-256 is:

c5c00ae774eccd17c39bd4da601c5ab0605af699bd2970b6cf3f3eb06987bbc3

These identifiers support traceability between the published evidence metadata and the corresponding internally retained artifacts.

The final checkpoint digest is reported from the checkpoint metadata. The complete proprietary checkpoint was not independently rehashed for this public publication, and the public evidence has not yet undergone independent third-party verification.

3. Contents of This Evidence Release

This Zenodo version contains a documentation-only evidence archive comprising:

• A technical README documenting the 8B source-token milestone.

• A machine-readable JSON status record containing milestone counters, identifiers, checksums, and governance status.

• Zenodo version notes explaining the release's scope and relationship to the original R1 evidence.

• SHA-256 checksums supporting verification of the publicly distributed documentation files.

The package is intended to enable transparent examination of the documented development milestone without disclosing proprietary training assets.

4. Model Lineage and Historical Evidence

This publication supplements the existing Agastya Native Brain evidence lineage, including previously published company-generated documentation of model initialization, tokenizer provenance, checkpoint identification, and native-model development.

The earlier historical evidence release remains independently accessible:

Original R1 DOI: https://doi.org/10.5281/zenodo.23054425

Concept DOI for the version family: https://doi.org/10.5281/zenodo.23054424

The present R2 publication is an additional evidence version and does not overwrite, invalidate, or retroactively modify the original R1 archive.

5. Evaluation, Acceptance and Production Boundaries

The completion of the 8-billion-source-token training milestone establishes a recorded candidate-training completion event. It does not, by itself, establish that the candidate has passed all subsequent evaluation, alignment, safety, independent reproducibility, or production-deployment gates.

At the time of this evidence release:

• Candidate-training milestone: Recorded complete

• Production promotion authorization: Not authorized

• Independent third-party evidence verification: Pending

• Same-suite external benchmark qualification: Pending

• Final production acceptance: Not claimed

A separate subsequent base-repair canary remained under rejection/hold and is not represented as an accepted production candidate in this release.

This publication makes no claim that Agastya has achieved independently verified artificial general intelligence, global benchmark superiority, or an unqualified national or international first.

6. Intellectual Property, Licensing and Access

Copyright © 2026 Rabbit Industries Pvt Ltd.

The publicly deposited evidence documentation and associated metadata are provided under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence.

This licence applies only to the public evidence materials included in this deposit.

The following proprietary materials are not distributed or licensed through this release:

• Trained model-weight binaries and private checkpoints

• Raw training corpora and underlying datasets

• Complete tokenizer vocabulary and merge artifacts

• Proprietary architecture and training source code

• Private infrastructure configuration, credentials, or operational secrets

The publication therefore represents a public technical evidence archive, not an unrestricted model-weight release or a production deployment authorization.

7. Public Evidence and Verification Links

GitHub 8B Evidence Release:
https://github.com/RabbitIndustries/agastya-native-brain-evidence/releases/tag/v1.1.0-f3-8b-token-evidence-r1

Hugging Face Evidence Repository:
https://huggingface.co/RabbitIndustries/Agastya-AGI-v01-Evidence

Official ZIPPRI Evidence Portal:
https://zippri.com/agastya-native-brain-evidence/f3-8b/

Original Zenodo R1 Evidence:
https://doi.org/10.5281/zenodo.23054425

8. Conclusion

Rabbit Industries Pvt Ltd is publishing this versioned evidence record to establish a dated and technically traceable account of Agastya F3 R2's reported 8-billion-source-token candidate-training completion.

The record provides auditable documentation identifiers and a persistent archival reference while clearly distinguishing training completion from final model acceptance, independent scientific verification, and production authorization.

Further validation results, externally verified evaluations, and additional development milestones may be documented through separate, appropriately versioned evidence releases.