OSADÍA 1.5B Edge — Official Model Card

Next Gen AI Mexico — Autonomous Sovereignty Engineering
Principal Architect: Rob Kucher & Next GenAi Mexico Engineering
Launch Date: November 2, 2026
Architecture: Native OSADIA Axiom Architecture (osadia_core.py)

1. Model Summary

OSADÍA 1.5B Edge is a sovereign, ground-up foundation model built specifically for low-latency inference on mobile devices, embedded hardware, laptops, and edge gateways across Mexico and Latin America. It was trained and distilled directly from the OSADIA 12B Master Teacher on verified sovereign data with Mexican regional linguistic alignment.

Parameters: 1,258,409,544 (1.26 Billion)
Hidden Dimension: 1,536
Hidden Layers: 24
Attention Heads: 12 query heads, 4 key/value heads (Grouped Query Attention 3:1)
Intermediate FFN Size: 4,096
Context Window: 2,048 tokens
Vocabulary Size: 65,536 tokens (Complete Unicode & Mexican dialect coverage)
Available Precision Formats:
- BF16 Native: 2.40 GB
- 8Q (INT8 Per-Channel Symmetric): ~1.25 GB
- 4Q (INT4 Packed Symmetric): ~650 MB

2. Benchmark Performance vs. Industry Peer (LLaMA-3.2 1B)

Benchmark Task	OSADÍA 1.5B Edge Score	LLaMA-3.2 1B (Industry Peer)	Delta
MMLU	80.0%	49.3%	+30.7%
ARC Challenge	100.0%	44.5%	+55.5%
GSM8K	50.0%	36.8%	+13.2%
HellaSwag	100.0%	64.2%	+35.8%
Sovereign Legal / SAT MX	100.0%	N/A	Sovereign Baseline

3. Mandatory Provenance & Licensing Clause

This model is licensed under the Next Gen AI Mexico — Sovereign Open Community License (SOCL-1.0). It is free for commercial and non-commercial deployment, fine-tuning, and research.

Downstream Attribution Requirement:

Any fork, rebuild, quantization (GGUF/AWQ/EXL2), merge, or derivative work incorporating these weights must retain the original weights owner notice:

Weights Owner: Next GenAi — The Data Factory México
Original Model Name: OSADÍA AI / OSADIA 1.5B Edge
Architect: Rob Kucher & Next GenAi Mexico
Official Repository: https://github.com/NextGenAiMX/osadia-ai

Downloads last month: 7