YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Shashwat Mudgal — Foundation Model Architectures
This organization hosts a family of research-driven language model architectures designed for different cognitive roles within modern AI systems.
Rather than producing a single monolithic model, this work separates reasoning, expression, and representation into distinct but interoperable architectures.
Model Families
Indra — Project Timeline
Indra is an independent Indic-language model architecture research project developed by Shashwat Mudgal (Divyansh Bharadwaj), in active development since October 2025.
The project has progressed through several architectural generations:
- Indra1 — dense transformer (GQA, YaRN RoPE), ~4B parameters
- Indra2 — MLA + MoE + sliding-window attention, ~500M parameters
- Indra3 — depth-recurrent architecture with adaptive computation time (ACT) halting and multi-head latent attention (MLA)
- Indra4 — hybrid architecture combining Gated DeltaNet, MLA+DSA attention, mixture-of-experts, ACT depth-recurrence, and multi-token prediction
- Indra-Lite — smaller dense models for constrained-hardware training
Some checkpoints were trained from scratch; others were initialized via weight-transfer from existing open-weight models and then substantially modified.
The architecture implementations, training code, and methodology for this project are not publicly released at this time.
This page exists to establish a public, dated record of the Indra project's development timeline.
🧠 INDRA
Integrated Neural Dual-attention Reasoning Architecture
INDRA is a reasoning-first decoder architecture, designed for:
- Stable logical generation
- Long-context reasoning
- Architectural experimentation (MoE, memory, routing)
- Deterministic and inspectable behavior
INDRA models are:
- Architecture-first
- Not chat-tuned
- Not roleplay-oriented
- Intended as reasoning backbones
Current base:
- INDRA (Qwen3-4B initialization)
🎭 URVASHI
Expressive Language Architecture
URVASHI is an expressive, personality-capable architecture, designed for:
- Roleplay and character modeling
- Emotional and stylistic diversity
- Model merging and synthesis
- Instruction-following and creative tasks
URVASHI models are:
- Merge-friendly by design
- Expressive rather than strictly logical
- Intended to absorb many fine-tuned variants
Planned variants include:
urvashi_llamaurvashi_rpurvashi_codeurvashi_dolphin- Composite merged URVASHI bases
🧩 VARUNA
Vector Alignment & Relational Understanding Neural Architecture
VARUNA is a representation and embedding architecture, designed for:
- Semantic embeddings
- Retrieval and clustering
- Alignment and similarity search
- Use alongside INDRA or URVASHI
VARUNA models are:
- Encoder-based
- Pooling-driven
- Task-agnostic
- Optimized for embedding quality, not generation
Design Philosophy
Each architecture serves a distinct cognitive role:
| Architecture | Role |
|---|---|
| INDRA | Reasoning & structure |
| URVASHI | Expression & personality |
| VARUNA | Representation & alignment |
They may be used independently or combined in larger systems.
Licensing Philosophy
- INDRA: Research-first, commercial-gated
- URVASHI: Expressive research, merge-friendly, commercial-gated
- VARUNA: Intended for broader embedding research (license may vary)
See individual repositories for full license terms.
Author
Shashwat Mudgal
Independent Researcher
Focus: Architecture-first language models
Disclaimer
These architectures are independent research efforts. Any initialization from existing models occurs at the weight level only and does not imply design inheritance.