🧪 DGPL Experimental 1 (dgpl-experimental-1)
1.12B Total Capacity · 144.71M Active Parameters per Token · Native 1-Million Token Context · Single-Model LoopSpec Speculative Edge Acceleration
[!WARNING]
⚠️ EXPERIMENTAL RESEARCH PROTOTYPE · NOT FOR PRODUCTION CRITICAL USE
dgpl-experimental-1is an early-stage experimental foundation research prototype developed by Durbhasi Gurukulam Private Limited (DGPL) to explore the frontiers of ultra-lightweight edge Mixture-of-Experts (MoE), recurrent looped reasoning (16 physical $\to$ 48 effective layers), and single-model LoopSpec self-speculative decoding on consumer edge hardware.Known Experimental Boundaries & Research Observations:
- Strengths: High execution throughput (61–85 tok/s on NVIDIA RTX 2050 4GB / >220 tok/s CPU), strong general logical deduction, basic arithmetic, physics derivations, and non-blocking Linux administration.
- Boundary Limitations: In complex non-homogeneous differential equations with resonance (e.g. $y'' - 4y' + 13y = e^{2x}\cos 3x$), the prototype may omit exponential multipliers in trial ansatzes ($y_p$) or enter repetitive loops without sampling penalties.
- Sampling Hardening Requirement: To avoid n-gram repetition loops during long reasoning sequences, inference engines MUST enforce non-zero penalties:
repeat_penalty: 1.25,presence_penalty: 0.5,frequency_penalty: 0.4,repeat_last_n: 256.
🏛️ Model Architecture Specifications
┌────────────────────────────────────────────────────────────────────────────────────────┐
│ DGPL EXPERIMENTAL 1 ARCHITECTURE SUMMARY │
├───────────────────────────────┬────────────────────────────────────────────────────────┤
│ Model Identifier │ dgpl-experimental-1 │
│ Total Parameter Capacity │ 1,123,553,280 Parameters (1.12 Billion) │
│ Active Compute per Token │ 144,711,680 Parameters (144.71M Active Compute) │
│ MoE Topology │ 1 Shared Expert + 15 Routed SwiGLU Experts (Top-2) │
│ Physical Layers │ 16 Transformer Layers │
│ Effective Recurrent Depth │ Up to 48 Effective Reasoning Passes (K=3 Looped) │
│ Attention Mechanism │ 14 Sliding Window (W=1024) + 2 Differential Attention │
│ Differential Attention Math │ Softmax(Q1 K1^T / sqrt(d)) - lambda * Softmax(Q2 K2^T) │
│ Native Context Window │ 1,048,576 Tokens (1M Tokens Default) │
│ 1M Context KV RAM Footprint │ < 272 MB RAM (Clamped Sliding Window Ring Buffer) │
│ Multi-Token Prediction (MTP) │ k = 1 Single-Model LoopSpec Speculative Decoding │
│ Empirical GPU Speed │ 61.6 – 85.4 tok/s (NVIDIA GeForce RTX 2050 4GB) │
└───────────────────────────────┴────────────────────────────────────────────────────────┘
📦 5-Tier GGUF Quantization Matrix
| GGUF Artifact | Quantization Format | File Size | Recommended Deployment Target |
|---|---|---|---|
dgpl_experimental_1_f16.gguf | Full Precision FP16 | 1,080.50 MB | High-VRAM GPU Serving & Baseline Research |
dgpl_experimental_1_q8_0.gguf | 8-Bit Quantized | 572.30 MB | High-Fidelity Desktop & Server Serving |
dgpl_experimental_1_q6_k.gguf | 6-Bit K-Quant | 428.15 MB | Balanced Precision for Standard 4GB–6GB GPUs |
dgpl_experimental_1_q5_k_m.gguf | 5-Bit Medium | 374.80 MB | 4GB GPUs & Compact Laptop Accelerators |
dgpl_experimental_1_q4_k_m.gguf | 4-Bit Edge Medium | 312.45 MB | <350 MB VRAM · NVIDIA RTX 2050 / 3050 & Edge Devices |
🚀 Quickstart: Local Serving via Ollama
1. Create Modelfile
FROM ./dgpl_experimental_1_q4_k_m.gguf
TEMPLATE """<|im_start|>system
You are an expert reasoning engine. Solve the given problem step-by-step.<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
<think>
"""
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER stop "</think>"
PARAMETER temperature 0.2
PARAMETER top_p 0.95
PARAMETER repeat_penalty 1.25
PARAMETER presence_penalty 0.5
PARAMETER frequency_penalty 0.4
PARAMETER repeat_last_n 256
PARAMETER num_ctx 32768
PARAMETER num_predict 4096
PARAMETER num_gpu 99
2. Build and Run
ollama create dgpl-experimental-1:1.1b -f Modelfile
ollama run dgpl-experimental-1:1.1b "Calculate 47 * 89 step by step."
🏢 Sovereign Corporate Attribution
- Entity Legal Name: Durbhasi Gurukulam Private Limited (DGPL)
- Registered Office:
2., P. NO. 18, Radha Vihar, Manchwa, Jaipur, Rajasthan 303706, Bharat - Official Support:
support@durbhasigurukulam.com|+91 7852034945 - Governance: Sovereign DGPL Edge AI Research Infrastructure
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
© 2026 Durbhasi Gurukulam Private Limited (DGPL). All rights reserved.