SAGI - Swarm AGI Language Model

SAGI is a novel causal language model that integrates swarm intelligence dynamics with transformer architecture. The model treats cognition as a dynamic, adaptive system where multiple internal "agents" collaborate through differentiable routing, trust mechanisms, and shared memory.

Model Description

Property	Value
Parameters	52.72M
Architecture	Transformer Decoder + Swarm Dynamics
Hidden Size	512
Layers	6
Attention Heads	8
Context Length	2048
Vocabulary	GPT-2 tokenizer (50,257 tokens)

Key Innovations

• Differentiable Routing: Continuous mixture-of-experts via attention (DiffRouter) instead of hard module selection
• Adaptive Gating & Trust: MetaController activates capacity under resource constraints; trust dynamics bias reliable components
• Episodic + Semantic Memory: Dual memory system with trainable retrieval utility
• Curiosity Engine: Injects novel goals when surprise is low, promoting exploration
• Self-Model & Rollback: Predicts state transitions and detects anomalies for self-correction
• Resource Dynamics: Soft conservation with learned converter; cognition consumes/recovers compute, memory, energy
• Value Monitoring: Tracks alignment to core values and freezes plasticity under drift

How It Works

┌─────────────────────────────────────────────────────────┐
│                       SAGI Model                         │
├─────────────────────────────────────────────────────────┤
│  ┌─────────────────┐      ┌─────────────────────────┐   │
│  │   Swarm-7 V2.2  │─────▶│  Swarm State S, T       │   │
│  │  (Cognitive     │      │  (Working Memory)       │   │
│  │   Dynamics)     │      └───────────┬─────────────┘   │
│  └────────▲────────┘                  │                 │
│           │                           ▼                 │
│           │              ┌─────────────────────────┐    │
│           │              │  Transformer Decoder    │    │
│           │              │  - Swarm-conditioned    │    │
│           │              │    attention & FFN      │    │
│           │              │  - RoPE embeddings      │    │
│           │              └───────────┬─────────────┘    │
│           │                          │                  │
│  ┌────────┴────────┐      ┌─────────────────────────┐   │
│  │   Observation   │◀─────│      LM Head            │   │
│  │   (from tokens) │      └─────────────────────────┘   │
│  └─────────────────┘                                    │
└─────────────────────────────────────────────────────────┘

The swarm processes observations derived from token embeddings, updating its internal state S. This state conditions the transformer's attention patterns and feed-forward activations via learned projections, creating bidirectional information flow between symbolic (tokens) and subsymbolic (swarm dynamics) processing.

Usage

Installation

pip install torch transformers datasets

Quick Start

from transformers import AutoTokenizer
from transformers import  AutoModelForCausalLM, AutoConfig

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained("reaperdoesntknow/SAGI")
tokenizer = AutoTokenizer.from_pretrained("reaperdoesntknow/SAGI")

# Generate text
model.eval()

prompt = "Once upon a time"
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=100,
    temperature=0.8,
    top_k=50,
    top_p=0.9,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Model Architecture Details

Swarm Configuration

Parameter	Value	Description
max_agents	20	Number of internal cognitive agents
dim_s	64	State dimension
dim_t	32	Task/goal dimension
dim_obs	48	Observation dimension
topk_route	5	Sparse routing top-k
K_thought_max	5	Maximum thinking iterations per step

Resource Budgets

Resource	Budget	Description
Compute	60.0	Compute budget per step
Memory	20.0	Memory capacity
Energy	25.0	Energy budget

Trust & Plasticity

• Trust Learning Rate: 0.07
• Fast EMA (Plasticity): 0.10
• Slow EMA (Consolidation): 0.002
• Core Values: ["truth", "safety", "efficiency"]

Limitations

• Early Research Model: This is an experimental architecture exploring swarm-transformer integration
• Training Data: Currently trained on TinyStories subset; may produce simple, story-like outputs
• Compute Requirements: Swarm dynamics add overhead compared to standard transformers
• Generation Quality: Model is undertrained; outputs may be repetitive or incoherent

Intended Use

This model is intended for:

• Research into multi-agent cognitive architectures
• Exploration of dynamic, adaptive language models
• Educational purposes in understanding swarm intelligence + LLMs

Not intended for:

• Production applications
• Safety-critical systems
• Generation of factual content

Training Details

• Dataset: TinyStories (subset)
• Optimizer: AdamW (lr=3e-4, betas=(0.9, 0.999), weight_decay=0.01)
• Scheduler: Cosine annealing
• Precision: FP32
• Hardware: CPU training (compatible with CUDA)

Citation

@software{sagi2026,
  title={SAGI: Swarm AGI Language Model},
  author={Reaperdoesntknow},
  year={2026},
  url={https://huggingface.co/your-reaperdoesntknow/SAGI}
}

---

Convergent Intelligence Portfolio

Part of the Standalone Models by Convergent Intelligence LLC: Research Division

Mathematical Foundations: Discrepancy Calculus (DISC)

SAGI's swarm intelligence dynamics connect to Discrepancy Calculus through Discrepancy Mechanics (Ch. 16 of the DISC monograph) — a reformulation of dynamics that replaces Newton/Lagrange with four discrepancy laws:

• DL0 (Co-Motion): Agent kinematics via metric derivative and environment flow
• DL1 (Discrepancy Energy): $E_{\text{disc}}[f] = \frac{1}{2}\int w(x)(Df(x))^2 d\mu(x)$ — stability through bounded discrepancy energy
• DL2 (Force as Discrepancy Gradient): Trust routing gradients as Euler-Lagrange from discrepancy action
• DL3 (Reciprocity): Symplectic invariance preserved across agent interactions

The discrepancy operator $Df(x) = \lim_{\varepsilon \downarrow 0} \frac{1}{\varepsilon} \int_x^{x+\varepsilon} \frac{|f(t) - f(x)|}{|t - x|} dt$ quantifies the local mismatch in each agent's contribution. The trust mechanism between agents is operationally a discrepancy energy minimization: agents whose outputs have high mutual discrepancy are weighted down; agents converging on shared structure are amplified.

Classical mechanics is recovered as a degenerate smooth limit of Discrepancy Mechanics — just as standard single-head attention is a degenerate limit of swarm routing.

Full theory: "On the Formal Analysis of Discrepancy Calculus" (Colca, 2026; Convergent Intelligence LLC: Research Division).

Related Models

Model	Downloads	Format
SMOLM2Prover	56	HF
SMOLM2Prover-GGUF	150	GGUF
DeepReasoning_1R	16	HF
S-AGI	0	HF

Top Models from Our Lab

Model	Downloads
Qwen3-1.7B-Thinking-Distil	501
LFM2.5-1.2B-Distilled-SFT	342
Qwen3-1.7B-Coder-Distilled-SFT	302
Qwen3-0.6B-Distilled-30B-A3B-Thinking-SFT-GGUF	203
Qwen3-1.7B-Coder-Distilled-SFT-GGUF	194

Total Portfolio: 49 models, 22,598 total downloads

Last updated: 2026-03-28 12:58 UTC

---

From the Convergent Intelligence Portfolio

DistilQwen Collection — Our only BF16 series. Proof-weighted distillation from Qwen3-30B-A3B → 1.7B and 0.6B on H100. Three teacher variants (Instruct, Thinking, Coder), nine models, 2,788 combined downloads. The rest of the portfolio proves structure beats scale on CPU. This collection shows what happens when you give the methodology real hardware.

Top model: Qwen3-1.7B-Coder-Distilled-SFT — 508 downloads

Full methodology: Structure Over Scale (DOI: 10.57967/hf/8165)

Convergent Intelligence LLC: Research Division