LatitudeGames/Harbinger-24B

Harbinger-24B is a 24 billion parameter instruction-tuned model developed by LatitudeGames, built upon Mistral Small 3.1 Instruct. It is specifically designed for immersive text adventures and roleplay, focusing on enhanced instruction following, improved mid-sequence continuation, and strengthened narrative coherence over long sequences. Utilizing Direct Preference Optimization (DPO) techniques, Harbinger-24B produces polished outputs with fewer common AI artifacts, excelling in generating engaging and consistent storytelling.