Indie Devs Seek Local LLMs Under 20B for Real-Time, Low-Latency NPC Dialogue

Impossible-Skirt-803 · reddit · 2026-08-12

An indie game developer shared their experience and challenges with integrating AI-driven NPC dialogue systems. The team initially used cloud-hosted APIs, but found that network jitter and high Time To First Token (TTFT)—such as waiting 5 seconds for a response—severely degraded the player experience.

To achieve natural human speech cadence, the team set a strict response time target of under 400ms and decided to pivot to on-device local deployment. Balancing VRAM allocation with game engine assets, they are currently seeking recommendations for low-latency open-weight models under 20B parameters.

Original post →

More from coding & agent

coding & agent channel →