Mini toolsOverview
Locai L1-Large is an open instruction-tuned model built on Qwen3-235B Instruct, trained with Locai’s Forget-Me-Not framework to boost alignment while retaining math and science strength. It reports state-of-the-art Arena Hard v2 results, safer behavior on AgentHarm, and ships under Apache-2.0 with easy Transformers and vLLM use.
Description
Locai L1-Large fine-tunes Qwen3-235B Instruct using Parameter-Efficient LoRA plus an experience-replay and self-improvement recipe that aims to raise helpfulness and safety without catastrophic forgetting. The model card reports top Arena Hard v2 scores versus non-reasoning frontier baselines and a 17 percent AgentHarm improvement compared to the base model, while keeping strong mathematical and scientific reasoning. It targets practical deployment with Transformers and vLLM examples, was trained on a single 8× H200 node, and highlights low-resource language gains along with renewable-energy hosting. License is Apache-2.0, region listing is EU, and a technical report is linked for more details.
About Locai Labs
Locai Labs is a UK-based foundational AI company focused on building transparent, community-driven AI models and services, with a mission to empower users rather than just chase AGI.
Website:
locailabs.com
Related Models
Last updated: November 13, 2025
