🦁 Lionsgate AI Lab

Independent AI research lab exploring local LLM inference, model quantization, agentic systems, and software architecture.

We explore how open-weight language models behave in practical environments, with a focus on reproducibility, reliability, and the trade-offs involved in running advanced AI systems on local hardware.

What we explore

• Local inference & quantization — deploying, optimizing, and evaluating open-weight models on NVIDIA DGX Spark and other local hardware.
• Model validation — investigating numerical and behavioral regressions, reproducing reported issues, and independently verifying corrections.
• Agentic systems — evaluating coding agents, tool use, context management, and end-to-end task completion.
• Software architecture — designing dependable orchestration, execution boundaries, and agentic workflows.

Published research

MiMo-V2.6-Flash-MOPD · EXL3 2.52 bpw

A locally quantized checkpoint with deployment information, evaluation evidence, and an independently validated layer-8 correction published in October 2026, incorporating a community contribution by emkey1.

Our approach

We document experimental configurations, hardware, serving runtimes, limitations, and independently verified outcomes.

We value reproducible evidence over headline benchmark numbers, and welcome constructive technical feedback and contributions.

Evidence over hype. Reproducibility over anecdotes.

---

GitHub — Source code & research · Contact