Moxin LM: From SOTA Research to Efficient Deployment

• Open Creation: The Moxin-7B series is our truly open, SOTA-performing LLM and VLM. We build, fine-tune, and openly release our own models.

• Efficient Deployment: We specialize in extreme quantization, creating resource-efficient variants of popular models (like DeepSeek and Kimi) to run anywhere.

We unleash the power of reproducible AI 🚀. Explore our models below and on GitHub, and read our research on Moxin 7B (Open Creation) and MoE Compression (Efficient Deployment).

World Models & Embodied AI

Moxin’s open AI ecosystem extends from digital intelligence into the physical world through the NU World Model & Embodied AI research initiative.

We build and openly release world models, physical-reasoning benchmarks, embodied agents, datasets, and evaluators for systems that simulate, predict, and act in the physical world.

• PhyGround — a human-annotated benchmark for evaluating the physical plausibility of generated videos, accompanied by the open PhyJudge-9B evaluator.
• PhyWorld — a physics-aware video-generation world model post-trained for temporal coherence and physical plausibility using preference data from PhyGround.
• Flash-WAM — a modality-aware distillation framework for accelerating joint video-action world models while preserving embodied-task performance.

Explore the research on GitHub · Browse all models and datasets · Visit Moxin