π² DeepNarra AI Research & Engineering
Sovereign Neural Intelligence Built from First Principles
Continuous Latent Reasoning β’ Gated Linear Attention β’ Speculative Leap Decoding β’ Two-Tier Hierarchical MoE
ποΈ Hugging Face Fleet β’ π Open-Source Governance β’ β‘ Model Matrix
ποΈ About DeepNarra
DeepNarra is an independent AI research organization dedicated to developing sovereign, high-efficiency neural architectures designed for extreme reasoning depth, deterministic stability, and high-throughput on-device inference.
DeepNarra unifies breakthroughs across Symplectic Hamiltonian continuous latent spaces, Chunk-wise Gated Linear Attention (GLA) recurrence, DAG-guided speculative leap drafting, and 1.58-bit ternary quantization (TQ1_0). Every architecture is validated for zero numerical degradation, transparent dataset lineage, and open-source compliance under the Apache License 2.0.
π The DeepNarra Fleet Matrix
| Model | Designation | HF Stored Parameters | Active / Token | Primary Architectural Invariant | Target Latency / Throughput |
|---|---|---|---|---|---|
π DeepNarra/Paham |
The Sovereign Sage | 4,171,887 (4.17M) |
Adaptive | Shannon Entropy 3-Gear Elastic Router (~1.2M β ~2.5M β ~4.17M), 75% GLA + 25% GLM-5 DSA, Zero-Drift Awake GGUF-QLoRA sidecars |
1,680+ tok/s |
π§ͺ Paham-Experimental |
3-Tier Continuum Sage | 4.17M + Adapters |
Adaptive | 3-Tier Awake-Sleep Continual Learning: Tier 1 Fast-Weights ($S_t$) + Tier 2 Sidecar-QLoRA ($r=8$) + Tier 3 MoE Router Gating | Experimental |
π DeepNarra/Hiraya |
The Sovereign Fruit & Flagship | 28,520,370 (28.5M) |
6.2M |
768-D Coconut Continuous Latent Space (<think>), Symplectic StΓΆrmer-Verlet dynamics, Minkowski Lorentz Hyperboloid |
378.5 tok/s |
π² DeepNarra/Narra |
The H-MoE Bedrock Tree | 28,519,602 (28.5M) |
6.2M |
Strictly Locked 128K Context (131,072), Chunk-wise GLA Recurrence, Lyapunov Stability Bounding (6.0 B/rec state pool) |
325.2 tok/s |
π¦ DeepNarra/Mako |
The Speed Predator | 28,519,602 (28.5M) |
6.2M |
Multi-Token Prediction (k=4 MTP), Prompt Lookup Decoding (k=50 PLD), O(1) Parallel Causal DAG Attention Tree |
1,000β3,000+ tok/s draft |
π Hiraya-1-Reasoning |
Phase 4 Reasoning Summit | 908,645,372 (0.91B) |
~100M |
2,500 steps Autonomous Boxing Gym, SymPy Formal CAS verification, DeepSeek-R1 GRPO (+1.45 verifier reward) |
92 Sub-Experts |
π Hiraya-1-Instruct |
Phase 3 14-Horizon Agent | 908,645,372 (0.91B) |
~100M |
Aligned across 14 Horizons (Prose, Code, Math, CUA <click>, MUA <touch>) with DoReMi auto-curriculum & OHEM gate |
92 Sub-Experts |
β‘ Hiraya-1-H-MoE |
Phase 2 92-Expert H-MoE | 908,645,372 (0.91B) |
~100M |
23 Heterogeneous Macro-Engines Γ 4 Micro-Experts = 92 Sub-Experts per layer + Persistent Shared Foundation Expert | 736 Total Experts |
ποΈ Hiraya-1-Base |
Phase 1 Dense Foundation | 117,993,216 (118M) |
80.5M Tied |
75% Mamba-2 SSD + 25% Differential Attention + BitNet b1.58 ternary weights + SwiGLU | Muon + AdamW |
ποΈ Core Architectural Pillars
1. π Hiraya: Continuous Latent Phase Space (Coconut Deliberation)
Hiraya bypasses the discrete autoregressive token bottleneck by computing reasoning chains directly in a 768-dimensional continuous latent manifold. Deliberation steps evolve according to Symplectic StΓΆrmer-Verlet Hamiltonian physics ($\Delta E < 10^{-13}$ energy preservation) and Minkowski Lorentz Hyperboloid geometry, projecting reasoning trajectory states into natural language only when conclusions solidify.
2. π² Narra: Bedrock GLA Recurrence & Locked 128K Context
Narra enforces a strictly bounded 128K token context window (131,072) that operates with linear computational complexity. By leveraging Chunk-wise Gated Linear Attention (GLA) with Lyapunov stability decay bounding, Narra maintains constant memory footprints across sustained million-token reasoning epochs without recurrent state drift.
3. π¦ Mako: Speculative DAG Leap Engine
Mako eliminates serial generation latency by coupling Multi-Token Prediction (k=4) with high-speed Prompt Lookup Decoding (k=50). A specialized $O(1)$ parallel causal DAG attention verification mask validates candidate draft tokens simultaneously in a single forward pass, unlocking generation rates from 1,000 to over 3,000 tokens per second.
4. π Paham: The Sovereign Sage (Omniversal Synthesis)
Paham unites the continuous latent exploration of Hiraya, the linear recurrent bedrock of Narra, and the speculative speed of Mako. Governed by a Shannon Entropy 3-Gear Elastic Router, Paham dynamically scales its active capacity from Edge Nano (~1.2M) to Deep Reasoning (~4.17M), integrated with Awake Continuum GGUF-QLoRA sidecars ensuring $0.00%$ base model degradation.
π± Sovereign Edge Deployment (1.58-bit TQ1_0)
All primary edge models are distributed in native 1.58-bit Ternary Quantization (model-TQ1_0.gguf) and serialized PyTorch Safetensors:
- Storage Footprint:
0.81 MB(Paham) to5.85 MB(Hiraya / Narra / Mako). - RAM Overhead:
< 15 MBtotal working set memory. - Target Hardware: Raspberry Pi, consumer laptops, mobile edge devices, and embedded ARM microcontrollers.
βοΈ Governance & Integrity
- Zero Fake Scores Policy: DeepNarra strictly forbids synthetic, estimated, or fabricated benchmark scores. All reported figures represent direct empirical telemetry from Colab runtime environments and safetensors tensor inspections.
- Open-Source Licensing: All model weights, configuration scripts, and architectures are released under the Apache License 2.0.
- Upstream Attribution Compliance: All pretraining and fine-tuning datasets are formally credited under their upstream open-source licenses (
FineWeb-Edu,Aya Collection,CodeFeedback-Filtered, andNo-Robots).