🧠 J.A.R.V.I.S. Titan V15: 14.8B MoE — 120M Token Frontier SFT
Principal Investigator: Dhanesh
Architecture: DeepSeekMoE Style (8 Routed Experts + 1 Shared Expert) with M4 Tri-Brid Active Gating
Training Hardware: Kaggle TPU v5e-8 (128 GB HBM, 8 Cores, Pure SPMD FSDP)
Dataset: 121.5M Tokens Curated & Token-Packed (dhanesh-hf/jarvis-v10-rft-dataset)
⚡ Model Overview
Jarvis-Titan-V15-MoE-120M-SFT is the frontier instruction-tuned MoE artifact developed for the J.A.R.V.I.S. Titan autonomous reasoning program. It builds upon the calibrated M4 Tri-Brid Memory backbone (dhanesh-hf/Jarvis-Titan-M4-Merged), adapting all 28 layers of routed experts, shared experts, and routing gates across a dense 120M-token curriculum.
Core Architecture Specifications:
- Base Parameters: 14.8B total | Active Parameters per Token: ~3.2B
- Layers: 28 | Hidden Dimension: 3584 | Head Dim: 128
- Attention Layout: Zero-Copy GQA (28 Query Heads, 4 KV Heads)
- MoE Topology: 8 Routed Experts ($d_{\text{inter}}=4736$) + 1 Isolated Shared Expert ($d_{\text{inter}}=4736$)
- Top-K Routing: Top-2 Active Routed Experts + 1 Shared Expert (3 active experts per token)
- Context Handling: Bounded $O(1)$ memory consumption via M4 Tri-Brid mechanics ($W=2048$ SWA, 1024-token Salient Reservoir, and Test-Time Neural Memory $M_t$).
🔬 Curriculum Composition
This model was trained exclusively on verified, high-density reasoning sequences:
- Verified STEM Mathematics: 42.0M tokens (34.6%)
- Olympiad Mathematics: 36.5M tokens (30.0%)
- Executable Code Synthesis & Repair: 18.3M tokens (15.0%)
- DeepSeek R1 CoT Reasoning: 19.6M tokens (16.1%)
- Dual Agentic Tool Execution: 4.7M tokens (3.9%)
🛡️ License & Terms of Use
This model is governed by the J.A.R.V.I.S. Titan Proprietary Research License (JTRL-v1.0).
- Non-Commercial Academic & Evaluation Use Only.
- Strictly Prohibited: Commercial deployment, paid APIs, weights redistribution, and commercial model distillation.
- Please review the full
LICENSEfile for legal details.
Citation
@misc{jarvis_titan_v15_sft_2026,
author = {Dhanesh},
title = {J.A.R.V.I.S. Titan V15: 14.8B MoE 120M High-Density SFT},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/dhanesh-hf/Jarvis-Titan-V15-MoE-120M-SFT}}
}