Krafton Raon 1 Mini tested - Local LLM setup

Summary

The video tests Krafton AI’s Raon 1 Mini Preview GGUF, a model described as a 30-billion-parameter mixture-of-experts (MoE) model with about 4 billion active parameters. Krafton says it was trained from scratch and supports Korean dialogue and agent tasks. The reviewer evaluates it in English.

Local Setup and Performance

The model was tested on Ubuntu using llama.cpp with 32 GB of system RAM. Most tests focused on a 16 GB GPU, with additional results reported from an RTX 5090.

Generation speed was about 31 tokens per second on the RTX 2000 and just under 190 tokens per second on the RTX 5090. The reviewer found the MoE model fast for its size and noted that it used very few tokens in many coding tasks.

Capability Tests

• Context recall: At a 256k context, the model struggled when the target information was placed at the 25% depth, failing to retrieve it in two of three attempts. Results at other depths were mixed, and some successful outputs were noisy.
• Logic: It scored 39/48 (81%) overall. It answered easy and medium questions perfectly, scored 11/12 on hard questions, and 4/12 on expert questions.
• Python tasks: It passed 74 of 100 tasks. The reviewer also reports an 83% response rate and says 11 questions received no answer.
• Seating-planner web app: Core features worked, including assigning and swapping guests, searching and filtering, changing table names and capacity, and saving changes. Some interactions were awkward or incomplete, but the model fixed errors after repeated prompting. The task used 68% of context.
• Sand simulation: The model implemented sand, water, walls, and acid, then corrected an initial material-behavior bug after prompting. Water could still teleport through sand, which the reviewer says is common in models. The task used 19% of context.
• Dungeon crawler: The player’s position updated in the interface but not in the game area, making movement unusable. After eight rounds of attempted fixes, the task still failed and used 68% of context.
• Blender via MCP: The model created a basic staff- or lantern-like object and produced screenshots after prompting. The result was not especially impressive, but required only a few additional prompts and used 34% of context.
• Godot via MCP: The model attempted a more complex 3D game with moving platforms and collectible items. The project eventually compiled, but key interactions remained broken, including collecting keys and navigating reliably. Repeated requests for fixes did not resolve the issues, so the test was considered a failure.

Overall Assessment

The reviewer’s main positive finding is the model’s low token consumption, paired with fast responses on suitable hardware. Raon 1 Mini often made quick corrections when given specific feedback, and the seating planner was its strongest practical coding example.

However, performance was uneven, particularly with persistent bugs and complex game tasks. The reviewer concludes that Raon 1 Mini is not comparable in capability to Qwen 27B, but may be useful for appropriately scoped tasks—especially front-end work—where speed and token efficiency matter.

Main Speaker and Sources

• Main speaker: The presenter from Luke’s DevLab, who runs the local tests and evaluates the results.
• Model information: Krafton AI’s Raon 1 Mini model card on Hugging Face.
• Test configuration: The presenter’s experiments using Ubuntu and llama.cpp. The video directs viewers to Luke’s DevLab website for configurations and results.