INDRA 8B Base
Ownership and Copyright
INDRA is an original language-model architecture and software project created by Shashwat Muthkul, also working under the pen name Divyansh Bhardwaj, and developed/released through DarkWhite Productions.
Copyright © 2026 Shashwat Mudgal (pen name: Divyansh Bhardwaj), operating through DarkWhite Productions. All rights reserved.
DarkWhite Productions is the organization through which this project is developed, maintained, and distributed.
This repository is manually gated. Approval to access the repository does not transfer ownership of the INDRA architecture, source code, model weights, documentation, or other proprietary materials.
Unless expressly stated otherwise, the original INDRA materials in this repository may not be reproduced, redistributed, relicensed, commercially exploited, modified for redistribution, or incorporated into another distributed work without permission from the copyright owner.
Third-party materials remain subject to their respective licenses.
See "LICENSE" and "NOTICE" for the applicable terms.
Integrated Neural Dual-Attention Reasoning Architecture
INDRA 8B Base is an experimental INDRA language model from the development of the INDRA architecture by Shashwat Mudgal.
This model represents the INDRA1 architecture at approximately 8 billion parameters — a larger-scale counterpart to the 4B INDRA1 checkpoint.
Important: INDRA is an architecture, not a fine-tune. The initialization model is used strictly as a weight substrate, not as a wrapper or dependency.
The checkpoint was developed as an architecture and language-model research experiment and should be understood as a historical generation of INDRA rather than as the current INDRA architecture.
Status: Historical experimental research checkpoint
Model Details
| Property | Value |
|---|---|
| Model | INDRA 8B Base |
| Architecture | INDRA1 |
| Parameters | ~8B |
| Model type | Causal language model |
| Developer | Shashwat Mudgal |
| Status | Historical experimental checkpoint |
| Primary purpose | Architecture research |
About INDRA
INDRA — Integrated Neural Dual-Attention Reasoning Architecture is an architecture-first language-model research project developed by Shashwat Mudgal.
The project investigates alternative language-model architectures with an emphasis on reasoning, structured generation, long-context processing, and architectural experimentation.
INDRA is developed as an independent research effort.
INDRA1
This model represents the same generation of the INDRA architecture
as DarkWhiteProductions/Indra-4B-Base, scaled to approximately 8
billion parameters.
INDRA1 was an experimental architecture developed to investigate alternative approaches to the internal computation of a causal language model.
The architecture and implementation in this repository should not be assumed to be identical to later INDRA generations.
Initialization
This model was developed using pretrained Qwen weights as an initialization point.
The use of an upstream pretrained model as initialization should not be interpreted as making this an unmodified Qwen model. All weights are mapped into INDRA-native modules, not reused via wrapper layers.
The resulting checkpoint represents an experimental INDRA architecture and subsequent training work.
For the exact upstream model and applicable licensing terms, refer to the corresponding upstream Qwen model documentation and license.
What This Model Is / Is Not
✅ INDRA Is
- A native architecture implementation
- A research-grade base model
- A stable foundation for further training or modification
- Suitable for controlled experiments and architectural work
❌ INDRA Is Not
- Instruction-tuned
- RLHF-aligned
- Chat-optimized
- Behavior-conditioned
Intended Use
This model is intended for:
- Research into language-model architectures
- Study of the INDRA1 architecture at larger scale
- Experimental reasoning research
- Historical analysis of INDRA development
- Language-model experimentation
- Reproduction and analysis of the released checkpoint
Out-of-Scope Use
This model is not intended to be treated as:
- A production conversational assistant
- A safety-critical AI system
- A medical, legal, or financial advisor
- An autonomous decision-making system
- A state-of-the-art reasoning model
- The current INDRA architecture
This is a historical research checkpoint.
Capabilities
INDRA 8B Base is an autoregressive causal language model capable of generating text from textual context.
Its capabilities depend on the INDRA1 architecture, the upstream initialization, subsequent training, and the particular checkpoint represented by this repository.
The model should not be assumed to possess reliable reasoning capabilities solely from its architectural design or parameter count.
Limitations
As an experimental research checkpoint, the model may exhibit:
- Hallucination
- Incorrect reasoning
- Repetition
- Inconsistent generation
- Poor instruction following
- Factual inaccuracies
- Prompt sensitivity
- Context-related limitations
- Training artifacts
- Unexpected behavior resulting from experimental architecture and training
No claim of state-of-the-art performance is made unless supported by documented evaluation results.
Relationship to Other INDRA Models
This INDRA1 checkpoint predates later generations including INDRA2, INDRA3, INDRA4, and INDRA-Lite.
Later models may differ substantially in:
- Architecture
- Attention mechanisms
- Tokenization
- Initialization
- Training procedure
- Model scale
- Memory mechanisms
- Routing
- Long-context implementation
The existence of this checkpoint does not imply that later INDRA models use the same implementation.
Reproducibility
When conducting research using this checkpoint, record:
- Exact repository revision
- Model checkpoint
- Configuration
- Tokenizer
- Generation parameters
- Transformers version
- PyTorch version
- Hardware
- Training configuration
Results from later INDRA generations should not automatically be considered comparable with this model.
Safety Considerations
This is an experimental language model and has not been presented as a fully safety-aligned conversational system.
It may generate inaccurate, biased, offensive, or otherwise undesirable content.
Appropriate safeguards should be implemented before using the model in applications.
Licensing and Third-Party Components
This model incorporates pretrained upstream model weights as its initialization.
Users must comply with the applicable license governing the upstream Qwen model as well as the license governing this repository and its original INDRA components.
Users should review all applicable licensing terms before redistributing, modifying, or commercially deploying the model.
The INDRA architecture and original implementation are independent research work by Shashwat Mudgal.
The use of upstream initialization weights does not transfer ownership of the INDRA architecture to the upstream model authors.
This section is not legal advice.
Attribution
INDRA
INDRA — Integrated Neural Dual-Attention Reasoning Architecture
Developed by:
Shashwat Mudgal
Upstream Initialization
This model was developed using pretrained Qwen weights as an initialization point.
The applicable Qwen model should be acknowledged according to its upstream license and documentation.
Citation
If you use this checkpoint in research or derivative work, please identify the exact checkpoint and acknowledge the INDRA project.
Suggested citation:
Mudgal, Shashwat. INDRA 8B Base:
An INDRA1 Architecture Research Checkpoint. 2026.
Created by Shashwat Mudgal, under the pen name Divyansh Bhardwaj, and developed/released through DarkWhite Productions.
Disclaimer
INDRA is an independent research project developed by Shashwat Mudgal.
This model incorporates pretrained upstream weights for initialization.
The use of those weights does not imply affiliation with or endorsement by the upstream model authors.
INDRA1 and subsequent INDRA architectures are research projects and should be evaluated independently before deployment.
- Downloads last month
- 409