Long-context models often face a choice between caching attention state for every token and using architectures that can lose useful context. This preview takes a hybrid route: most layers use fixed-state linear attention, while a smaller set retains full attention. The result aims to make a 262K-token window practical without making every layer carry the same inference cost.
Key Capabilities
- Sparse computation: Six routed experts and one shared expert are active per token, using 4.3B of 28.8B total parameters. This reduces per-token computation relative to a dense model of similar total size, though it does not make the full model small to store.
- Hybrid long-context design: Eighteen linear-attention layers are paired with seven gated MLA layers. Only the latter keep a KV cache, so long prompts have a different memory profile from an all-full-attention model.
- Reasoning and tool use: The model supports thinking mode and function calling, with documented compatibility for agent harnesses including Hermes Agent, Pi, and OpenCode.
- Multilingual and Korean focus: Training includes Korean, English, Japanese, Chinese, Spanish, French, and Thai. KRAFTON reports a 92.9 score on its internal Korean conversation benchmark; that evaluation is not publicly available, so it is less independently verifiable than public benchmark results.
Who It Suits—and the Trade-Offs
A good fit for researchers and developers evaluating open-weight models for Korean conversation, long-context text tasks, or tool-using agents. Look elsewhere if you need image understanding: this model accepts text only. It is a preview, its repositories require an access request, and its headline benchmark results are vendor-reported. The downloadable model components use Apache 2.0; the hosted API and playground have separate terms.
