---
base_model: dhanesh-hf/Jarvis-Titan-V14-MoE-Merged
tags:
- deepseek-moe
- titans-neural-memory
- pallas-tpu
- ultra-long-context
- sliding-window-attention
- tri-brid
- jarvis-titan
- distillation
- reasoning
- agentic
library_name: transformers
pipeline_tag: text-generation
license: other
license_name: jtrl-v1.0
license_link: LICENSE
---


# J.A.R.V.I.S. TITAN 14.8B MoE — MILESTONE M3 ULTRA-LONG ADAPTER

Official Milestone M3 (Phase 3) weights for **J.A.R.V.I.S. Titan 14.8B DeepSeekMoE + Tri-Brid Memory Architecture**, distilled from full dense attention to Sliding Window Attention ($W=2048$) on Google Cloud TPU v5e-8.

## Distillation & Training Specifications
- **Base Model**: `dhanesh-hf/Jarvis-Titan-V14-MoE-Merged` (14.75B MoE, 100% Frozen)
- **Adapter Initialization**: `dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter` (Phase 2 Tri-Brid)
- **Dataset Source**: `dhanesh-hf/jarvis-v10-rft-dataset` (100% Real Non-NIAH Peer-Reviewed Papers & Code)
- **Sliding Window Size**: $W = 2048$ tokens (capping KV cache at ~115 MB)
- **Strategic Layers**: [3, 7, 11, 15, 19, 23, 27] (7 Memory Bridges)
- **Tier 2 Salient Reservoir**: $R=1024$ slots ($H_Q=28, H_{KV}=4$)
- **Tier 3 Titans Neural Memory**: $d=512$, Google Pallas TPU VMEM SRAM kernel
- **Triple-Gated Adaptive Fusion**: MAG-3 ($g_{\text{local}}, g_{\text{res}}, g_{\text{mem}}$)
- **Distillation Loss**: $(1 - 0.5) \mathcal{L}_{\text{CE}} + 0.5 T^2 \mathcal{D}_{\text{KL}}$ ($T=2.0$)
- **Total Adapter Parameters**: 90,044,458 (90.04M)
- **Total Tokens Trained**: 20,012,495