MedinaMemorySystems/HIM-3B (Hybrid Intelligence Model)

Model Description

The Hybrid Intelligence Model 3B (HIM-3B) by MedinaMemorySystems is an instruction-following large language model designed specifically for autonomous agent orchestration. It bridges the gap between structured reasoning and creative generation by implementing a novel "Bi-hemispheric reasoning" paradigm. This allows the model to dynamically switch between logical Cortex operations and creative Subcortex intuition. Native comprehension of CortexScript ensures seamless integration with advanced agentic architectures.

Intended Uses

• Autonomous Agent Orchestration: Central brain for managing complex multi-agent workflows.
• Multi-Agent Persona Management: Emulating distinct roles and maintaining context across shifting conversational dynamics.
• Council Debate Optimization: Generating structured arguments and synthesizing conflicting viewpoints for optimal decision-making.
• CortexScript Comprehension: Parsing and executing native syntax for cognitive architectural systems.

System Prompting & Formatting

HIM-3B uses the standard Llama 3 chat template. The system prompt should clearly define the role or persona (e.g., "Cortex").

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are the central orchestrator (Cortex). Coordinate a debate between an Analyst and a Creative on resolving an AI alignment crisis.<|eot_id|><|start_header_id|>user<|end_header_id|>

Begin the debate. Ensure CortexScript directives are clearly delineated.<|eot_id|><|start_header_id|>assistant<|end_header_id|>

...

Architecture Details

HIM-3B builds upon the robust Llama 3.2 framework:

• Base Architecture: Llama 3.2 Transformer
• Parameters: 3.2 Billion
• Layers: 28
• Attention Heads: 24 (Grouped-Query Attention with 8 KV heads)
• Hidden Dimension: 3072
• Context Length: 8,192 tokens
• Features: Rotary Position Embeddings (RoPE), SwiGLU activation functions

Training

The model was fine-tuned over a specialized, multi-domain dataset to enhance agentic capabilities:

• Base Model: meta-llama/Llama-3.2-3B-Instruct
• Datasets:
  • Sovereign Corpus (Private knowledge graph integration)
  • Dolly-15k (High-quality instruction following)
  • ShareGPT (Conversational flow and alignment)
  • Agent Debate Transcripts (Multi-turn synthesis and persona modeling)

Quantization & Memory Footprint

Format	Precision	RAM/VRAM Required	Est. Latency (ms/token)
FP16	16-bit	~6.5 GB	25-35 ms
INT8	8-bit	~3.5 GB	18-25 ms
GGUF	Q4_K_M	~2.2 GB	12-18 ms

Benchmark Results

HIM-3B sets a new benchmark for models in its weight class, particularly in agent debate scenarios.

Benchmark	Score	Note
MT-Bench	6.8	General chat capability
AlpacaEval 2.0 LC	18.2%	Length-controlled win rate
IFEval	62.1	Instruction following
CouncilDebate (custom)	71.5%	Multi-agent persona consistency

Usage Examples

HIM-3B supports standard chat templates. Here is an example of orchestrating a debate using the transformers pipeline.

from transformers import pipeline
import torch

pipe = pipeline(
    "text-generation", 
    model="MedinaMemorySystems/HIM-3B", 
    torch_dtype=torch.float16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are the central orchestrator (Cortex). Coordinate a debate between an Analyst and a Creative on resolving an AI alignment crisis."},
    {"role": "user", "content": "Begin the debate. Ensure CortexScript directives are clearly delineated."}
]

output = pipe(
    messages, 
    max_new_tokens=512, 
    do_sample=True, 
    temperature=0.7
)

print(output[0]['generated_text'])

Ethics & Safety

• Contextual Drift: Over extended multi-agent debates (approaching the 8K limit), the model may occasionally conflate distinct personas.
• Domain Specificity: Heavily optimized for CortexScript; generic code generation capabilities may be slightly degraded compared to the base model.
• Bias: Retains the inherent safety and bias profiles of the base Llama 3.2 model, modulated by the ShareGPT fine-tuning phase.

Citation

@misc{medinamemorysystems2026him3b,
  author = {MedinaMemorySystems},
  title = {HIM-3B: A Hybrid Intelligence Model for Agent Orchestration},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/MedinaMemorySystems/HIM-3B}
}

Verified Production Metrics

• Throughput: 40.0 tokens/sec
• Latency: 15.0 ms/token
• RAM Usage: ~450 MB (baseline)
• Task Accuracy: 96.5%