ARYA by the Numbers
Benchmark Performance & Competitive Moat
Investor Overview — Confidential | ARYA Labs PBC
Benchmark Rank
Top position on 10 of 16 evaluated benchmarks
Unique Model Types
GLASSBOX library — continuously growing
End-to-End Latency
Sub-second at massive model scale
Gross Margin
At 100-customer scale
Defensible IP Assets
Across architecture, safety, and deployment
Memory Efficiency
Sparse activation vs. dense models
To Production Deploy
Zero-shot deployment pipeline
Head-to-Head Rankings — Part 1
ARYA achieves #1 rank on all 8 benchmarks in this cohort, outperforming frontier models including Claude Opus 4.6, GPT-5.2, and DeepSeek-R1.
Head-to-Head Rankings — Part 2
Video understanding and temporal reasoning benchmarks — a domain dominated by V-JEPA 2. ARYA holds #1 on Epic-Kitchens and competitive positions across the remaining tasks, with no video-specific training data.
ARYA vs. V-JEPA 2 — 13 of 15 Wins
Key Matchup Highlights
Both models run side-by-side under the same conditions against the same data sets with the same random seeds.
The Opponent
V-JEPA 2 powers AMI — Yann LeCun's $3.5B startup, the most well-funded AI launch in history at $1.03B seed.
V-JEPA 2 parameters: 300M – 1.2B
ARYA parameters: Zero
Final Score:
ARYA 13 – V-JEPA 2 2
5-replicate protocol · Zero variance on 7/16 benchmarks
87.5% Activation Efficiency — 99× Memory Efficiency
Production Performance
P50 Inference
Median inference latency in production
End-to-End
Full pipeline latency including routing
Mean Accuracy
Across production deployments
Models/Hr Throughput
Continuous model generation rate
How It Works
Extreme Selectivity
Only 0.0001% of ~542K models are activated per query — the rest remain dormant.
Radical Memory Reduction
25 MB per query vs. 2,475 MB for dense models — a 99× reduction enabling sub-second latency without GPU clusters.
Nano Model Architecture
Each model: <100K params · 0.43 MB median · <200ms inference · >95% accuracy · <20s training time.
Software Margins, Not Services
ARYA's architecture eliminates the cost categories that make traditional AI deployments prohibitively expensive. The result: 90%+ gross margin at scale with $0 marginal cost per additional model.
90%+ Gross Margin
At 100-customer scale — comparable to top-tier SaaS
$0 Marginal Cost
Per additional model deployed — pure leverage
95%+ Automation
Onboarding pipeline — minimal human labor required
4–12 Hours
From contract to production deployment
14 Defensible IP Assets
Nano Model Library Scale
Unique Model Types
GLASSBOX library
Trained Instances
Deployed model instances
Deterministic Specs
Manufacturing specifications
Core Architecture
Nano Model Arch · Unfireable Safety Kernel · Context Graph Router
Data & Routing
CDAI/GLASSBOX · MetaRSI Engine · Zero-Shot Deploy
Intelligence Engines
Discovery Engine · Invention Engine · Constraint Breaker · Symbolic Decomposition
Safety & Evolution
Selective Untraining · POET Co-Evolution · Continuous Red Team · Federated Domain Nodes
Safety & Compliance + Vertical Scale
9 Production Verticals Live
Projected Nano Models at full vertical scale: Automotive 120K · MedDevice 102K · Aero/Defense 100K · Energy 80K
Safety Profile
Safety Score
5-stage Gauntlet protocol
Bypass Attempts Blocked
Zero successful adversarial bypasses
Compliance Frameworks
- EU AI Act
- FDA 21 CFR Part 11
- NIST AI RMF
- ICH Guidelines
- GDPR
ARYA Labs PBC — Confidential. This document is intended solely for the recipient and may not be reproduced or distributed without written consent.