Krafton launches Raon open-source AI model suite for speech and vision

Krafton’s Raon launch is a strong signal that top-tier gaming companies are no longer just AI customers—they’re becoming model developers in their own right. By shipping a 9B-parameter speech model, a low‑latency duplex speech agent, an open TTS system and a from‑scratch vision encoder, Krafton is building a full multimodal stack that can power in‑game NPCs, voice interfaces and beyond. The decision to open‑source both models and training data, rather than keep them proprietary, positions Raon as infrastructure for Korea’s broader AI ecosystem, not just a house tool for one publisher.

Strategically, Raon fits into a broader shift where entertainment and gaming firms see agents and real‑time multimodal interaction as their competitive moat in the age of commoditized text chatbots. Benchmarking directly against Google’s SigLIP2 hints that Krafton wants to be taken seriously as a foundational model lab, not just a content studio. For the global race to AGI, it’s another proof point that innovation is diffusing outward from a handful of US labs toward regional champions with deep domain data—here, decades of interactive media and game telemetry that can train highly contextual agents.