Krafton, the company behind PUBG: Battlegrounds and Battlegrounds Mobile India, has released Raon-1 nano, a small AI language model designed to run on a player’s own device instead of in the cloud. The model has an effective 1.5 billion parameters and is free to download and use commercially under the Apache 2.0 licence on Hugging Face. Krafton says it was built for natural conversation, following instructions and using tools, in both Korean and English.
The release matters for gamers because Krafton’s two best-known AI characters, PUBG Ally and inZOI’s Smart Zois, have so far run on small models supplied by Nvidia rather than on Krafton’s own.
KRAFTON AI, the company’s research arm, tied the model directly to its game work, calling the release the next step in “our journey from PUBG Ally toward on-device LLMs”. Kangwook Lee, Krafton’s chief AI officer, went further, saying the company “finally” has “a better base model for our edge AI applications”.
Edge AI is the industry term for models that run on local hardware, such as a gaming PC or a phone. Krafton has not said which of its games will use Raon-1 nano, or when.
PUBG Ally proved the idea, but it ran on Nvidia’s model
Krafton has been shipping on-device AI in its games for more than a year and a half. When inZOI entered early access in March 2025, its optional Smart Zoi feature turned the city’s residents into AI-driven characters, using a 0.5-billion-parameter Mistral NeMo Minitron model from Nvidia’s ACE platform, according to Nvidia. That feature ran only on GeForce RTX graphics cards. PUBG Ally followed the same route, and its AI teammate Ella reached players in the Ally Duo beta on Sanhok in June 2026.
Krafton explained the thinking behind Ally in an engineering interview published by Nvidia. Its team tested cloud-hosted models first, but found that network and inference delays together made responses feel too slow for live squad play. It switched to Nvidia’s 2-billion-parameter Mistral-NeMo-Minitron model, compressed it to fit beside the game in graphics memory, and got Ally running on cards with as little as 8GB of video memory. Cloud models would also have been expensive at scale, with Krafton’s own research putting Claude Opus 4.8 at $7.10 per replayed match.
Even the local version of Ally came with clear compromises for players. The Ally Duo beta was limited to PCs with GeForce RTX cards, and Krafton’s patch notes advised players with delayed responses to lower their graphics settings or resolution. In other words, the AI teammate and the game were competing for the same graphics memory. That is the problem any on-device model from Krafton would have to solve.
Raon-1 nano is built around the memory squeeze that held Ally back
Raon-1 nano looks designed with exactly that constraint in mind. Its 3.1 billion total parameters include a 1.5-billion-parameter table of per-layer embeddings, a technique Krafton is using for the first time “to reduce VRAM usage”, according to Lee. That leaves around 1.5 billion parameters doing the main work, which is smaller than the 2-billion-parameter model behind Ally. The model card also lists a context window of 262,144 tokens, which is long enough to hold a very long conversation or a large amount of game information.
Krafton built the model in an unusual way. It started from Raon-1 mini, a 30-billion-parameter mixture-of-experts model that it previewed in October. It then converted that into a smaller dense model using the method from its research paper, “Pruning and Distilling Mixture-of-Experts into Dense Language Models“. KRAFTON AI says the paper has been accepted as an oral presentation at NeurIPS 2026. The team then trained the model on a further 3 trillion tokens, before fine-tuning it for conversation, instruction following and tool use.
Krafton’s benchmark claims are strong, although not across the board. On its model card, Raon-1 nano leads similarly sized models such as Google’s Gemma 4 E2B and Alibaba’s Qwen3.5 2B on instruction following and Korean conversation, scoring 75.9 on IFBench. It trails on some tests, however, including the BFCL v4 tool-calling benchmark, where Liquid AI’s LFM3 scores 67.2 against its 56.3. Some of the Korean results also come from internal tests that have not been published.
Phone and Mac formats point beyond GeForce RTX
The release formats are the clearest hint of where Krafton could take its AI next. Smart Zoi and Ally both depended on GeForce RTX hardware, but Raon-1 nano ships in GGUF, MLX and LiteRT-LM versions, covering common PC tools, Apple silicon Macs and Google’s runtime for running models on phones.
The model card also lists “mobile control” among the tasks it was trained for. That opens the door to AI features on hardware that an RTX-only system could not reach.
Mobile is where most of Krafton’s players are. PUBG Mobile, BGMI and Peacekeeper Elite brought in just over half of the company’s revenue in its third-quarter results, and the studio is preparing lighter versions such as BGMI Air for lower-end phones. Krafton has not announced an AI teammate for mobile, and running even a small model next to a demanding shooter on a mid-range phone remains a hard problem. A free model that Krafton owns outright, and that already runs on phones, does remove one obstacle, since it no longer needs a third party’s model to experiment.
The model also fits Krafton’s wider AI push. The company launched its Raon model brand in April, with open models for speech, voice chat, text-to-speech and vision, and its research team has nine main-track papers at NeurIPS 2026, according to Korean reports. Lee said more releases are on the way, without giving dates. Raon-1 mini itself is still a preview, and Krafton has said an official version will follow with a full technical report.
