Ver todos os eventos

EmpoorioLM: Empoorio's Sovereign AI.

The first open weights language model of the Empoorio ecosystem. Hybrid Transformer+SSM, on-chain verifiable via zkML and designed to run from mobile to cluster.

25.4MParameters — small preset
Apache 2.0Open weights license
zkMLOn-chain verifiable
Phases 0–8Roadmap completed

TECHNICAL SPECIFICATIONS

Architecture, benchmarks and integration guide for EmpoorioLM.

Why EmpoorioLM
?

EmpoorioLM combines hybrid Transformer+SSM architecture with paged attention, multi-token prediction and SimPO alignment. It is the only language model whose inferences can be cryptographically verified on-chain via pallet-zk-verifier — no central custodian, no black box.

Ecosystem products<br/>powered by EmpoorioLM

ailoos

Personal assistant powered by EmpoorioLM. Local on-device processing with mini/small preset. No data collection.

params in local model
25M
params in local model
eoonia

Integrated AI Marketplace in Eoonia lets you hire EmpoorioLM inferences for TextGen, CodeGen and DataAnalysis directly from the wallet.

active subnet types
5
active subnet types
daotononos

On-chain governance assisted by EmpoorioLM. The model analyzes proposals, generates summaries and detects conflicts of interest automatically.

governance assistance
AI
governance assistance
kryptos

Operating system with integrated AI. EmpoorioLM acts as system assistant — digital identity management, credentials and on-chain keys in natural language.

on-chain verification
zkML
on-chain verification
Ailoos AI
Eoonia
DAOtonomos
KryptoOS

GPT-4 level architecture

  • 1

    Hybrid Transformer+SSM Architecture

    Mamba blocks (Selective SSM S6) interleaved with standard attention. Paged Attention with KVBlockTable and Copy-on-Write. Multi-Token Prediction for 1.8–2.4× faster decoding. FP8 training with half the VRAM of bfloat16.

  • 2

    Pre-training → SFT → SimPO Pipeline

    Pre-training on multilingual corpus (Don Quixote + seed + instructions). SFT from checkpoint. SimPO alignment without external reward model — 34+ annotated pairs. Distributed training via FSDP2 + Tensor Parallelism. Automatic loss spike detector.

  • 3

    Verifiable Inference via zkML

    Server compatible with /v1/completions (OpenAI API). Mini (4.3M, ~9 MB RAM) and small (25.4M, ~52 MB RAM) presets for edge and production. Inferences cryptographically verifiable on-chain via EmpoorioChain pallet-zk-verifier.

EmpoorioLM integration kit

Empoorio ecosystem tools and products that integrate EmpoorioLM as their AI engine.

EmpoorioLM — El Modelo de IA Soberano de Empoorio | empoorio