KRAFTON released and open-sourced Raon-1 mini Preview

AT A GLANCE

KRAFTON released and made available Raon-1 mini Preview weights, disclosing its architecture, evaluation results, and local deployment methods.

Article

PUBG developer open-sources 30B model: Bug fixing approaches Meta, with Kimi and DeepSeek technology under the hood

Dongcha Beating AI breaking news, the South Korean gaming giant KRAFTON has released Raon-1 mini Preview. This is a MoE language model trained from scratch, with approximately 28800000000 total parameters, activating only 4300000000 per inference, with a focus on optimizing Korean dialogue, programming, and Agent capabilities, and supporting 260000 Token context.

Although it was trained from scratch, its underlying technology also shows the influence of Chinese companies. It uses a hybrid attention architecture, and in its 25-layer network, 18 layers use a linear attention mechanism improved from Kimi Delta Attention. In addition, the team simplified DeepSeek's mHC hyperconnection technology to reduce training computational costs.KRAFTON tests released by Raon-1 show that SWE mini achieved 70.8%on the -Verified code-fixing test, surpassing Qwen3.6-35B-A3B's 63.2%and approaching 72.0%for Meta Muse Glimmer. However, in the Terminal-Bench Lite terminal-operation test, it scored only 47.0%, significantly behind Qwen3.6-35B-A3B's 63.3%and Muse Glimmer's 70.7%. In the IFBench instruction-following test, it reached 84.2%, higher than the two models mentioned above.

The model weights have been released under the Apache 2.0 license, with versions including GGUF、MLX available and support for local operation on Mac. The NVFP4 quantized version can be deployed on a single RTX 5090. The team also provides a DFlash2 speculative decoding model; in tests on a single NVIDIA B300, it achieved a maximum 2.18-fold increase in output throughput compared with standard decoding.

Original link https://m.theblockbeats.info/flash/373005

Key points

Raon-1 mini Preview is a KRAFTON language model trained from scratch and released by MoE.
The model has approximately 28800000000 total parameters, activating only 4300000000 per inference and supporting 260000 Token context.
The model is optimized primarily for Korean dialogue, programming, and Agent capabilities.
25-layer network contains 18 layers using a linear attention mechanism improved from Kimi Delta Attention.
The team simplified DeepSeek's mHC hyperconnection technology to reduce training computational costs.
KRAFTON Released tests show that Raon-1 mini achieved SWE on -Verified, 70.8%on Terminal-Bench Lite, and 47.0%on IFBench. 84.2%。
The model weights have been released under the Apache 2.0 license, with GGUF、MLX versions available and support for local operation on Mac.
NVFP4 The quantized version can be deployed on a single RTX 5090.
The team provides a DFlash2 speculative decoding model; in tests on a single NVIDIA B300, it achieved a maximum 2.18-fold increase in output throughput compared with standard decoding.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a MoE model that enables only a portion of its parameters during inference, so its total parameter count is approximately 28800000000, but only 4300000000 are activated each time. The report shows that it performs relatively well on code-fixing and instruction-following tests but is comparatively weaker on terminal-operation tests; open weights and multiple versions allow users to try local deployment.

Why it matters to readers

The preview also disclosed the model scale, long context, evaluation results, open license, and hardware deployment methods, providing developers with specific information for assessing its programming, Agent, and local-operation capabilities.

Beginners should focus on points: it is an open-weight MoE model; capabilities vary across different tests;GGUF、MLX and quantized versions may lower the barrier to local trial.

Risks and unknowns

The performance figures come from tests released by KRAFTON and the evidence contains no third-party reproduction or independent verification materials.
The evidence does not state the release date of the full version.
The evidence does not provide a weight repository address, so the availability of the files cannot be verified from this article alone.
The evidence does not specify the complete configuration, operating conditions, or comparison methods for each evaluation.

KRAFTON released and open-sourced Raon-1 mini Preview

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center