We build AI that loves humanity.
We research intelligence that understands, protects, and grows with human life, beyond systems that merely perform tasks.
Get in TouchResearch Areas
Humanity-Aligned AI
AI that loves humanity
We study AI that treats people not as users or data points, but as lives to understand, protect, and grow with. By love, we mean a technical orientation: prioritizing human dignity, vulnerability, agency, and flourishing.
- Human dignity as a design constraint
- Responses aware of vulnerability and risk
- Behavior evaluation and safety verification
- Interactions that expand agency and creativity
Adaptive Intelligence
Reasoning systems that learn and adapt
We research AI architectures that understand context, reason under uncertainty, and adapt efficiently at inference time. Our work turns into papers, patents, open models, and reproducible experiments.
- Inference-time learning and long-term memory
- Context understanding across sustained interactions
- MoE efficiency and adaptive computation
- Open models and reproducible evaluations
Research Output
Cross-Model Conversation Context Management System
Dec 18, 2025
Research on architectures for preserving conversation context and state across language models. It studies shared representations, format conversion, and session management so interactions with AI remain coherent as systems change.
Dynamic Computation Optimization for Language Models
Jan 15, 2026
MoE-based dynamic computation research for reducing inference cost and improving response efficiency in AI systems.
Inference-Efficient MoE via Non-Computational Experts
Mar 20, 2026
A modular optimization technique that extends the gating network of existing MoE models with non-computational experts — preserving the original model while substantially reducing inference cost.
Parameter Updater Expert in Mixture-of-Experts Layer
Mar 30, 2026
A neural architecture embedding a dedicated "updater expert" slot inside the MoE layer that dynamically modifies sibling experts' weights at inference. Enables the model to learn new rules at inference time without a separate training stage.
Publications
Parameter Updater Experts: Inference-Time Learning in MoE Models via DeltaNet-LoRA
Han, Jongyun · Apr 10, 2026
Proposes dedicated expert slots within MoE layers that generate weight modifications during forward passes. DeltaNet-LoRA achieves 100% in-context fact retrieval on OLMoE-1B-7B, and 80.1% persistent retrieval under sliding-window attention.
Available on: Zenodo opens in new tab · SSRN opens in new tab
Open Models
Nemotron-3-Super-64B-A12B-Math-REAP
Apr 22, 2026
REAP-pruned (512 → 256 experts, MTP layer removed) NVIDIA Nemotron-3-Super-120B-A12B, briefly LoRA-RL fine-tuned on AIMO3 + AstralMath-v1, then post-training quantized. Released in three variants for different deployment trade-offs.
AIME 2026 avg@4 — FP8: 0.9167 · AWQ: 0.9083 vs. 0.9000 base (120B).
Variant Model Hugging Face Links: BF16 opens in new tab · FP8 opens in new tab · AWQ opens in new tab
View benchmarks and details on GitHub → opens in new tab
Certified Venture Enterprise
Venture Company Certification
Max & Omnis has been certified as a venture enterprise under Korea's venture business framework, recognizing the company's innovation growth potential.
- Certificate No.
- 20260506030005
- Valid Period
- May 6, 2026 - May 5, 2029
About the Lab
Max & Omnis is an AI research lab building AI that loves humanity. We believe intelligence should not evolve toward replacing people, but toward understanding, protecting, and helping human life flourish. We turn that belief into testable research through patents, publications, and open models.
Contact
We welcome inquiries about research collaboration, publications, open models, and investment.
madmax0404@maxandomnis.com© 2026 Max & Omnis Inc. All rights reserved.