Krafton AI Open-Sources Raon-1 Mini: A 30B MoE with 4B Active Parameters Matching Muse Glimmer on Coding Autonomy

Background and the Problem

In gaming AI automation, autonomous NPC simulation, and local desktop coding agents, compact foundation models have reached an architectural inflection point. Conventional 7B–14B dense backbones frequently lack the parametric capacity to execute long-horizon multi-tool planning without goal drift. Conversely, dense 30B models impose heavy VRAM footprints and memory bandwidth demands that saturate consumer workstations (16GB–24GB VRAM like RTX 4090/5080). Sparse Mixture-of-Experts (MoE) architectures—pairing large parametric capacity with minimal per-token activation—offer an optimal path forward. Addressing this space, Krafton AI officially open-sourced Raon-1 mini (@Krafton_AI, Status #2112772568776339709) under Apache 2.0 on October 21, 2026.

Architecture and How It Works

1. 30B Total / 4B Active Sparse MoE Topology:

• Houses 30 billion total parameters while activating only 4 billion parameters per token forward pass;
• Delivers the rich world knowledge and syntactic comprehension of a 30B foundation model alongside the lightning execution speed of a 4B parameter model;

1. Dedicated DFlash Speculative Drafting:

• Released in tandem with a bespoke DFlash Draft Model tuned specifically to Raon-1 mini's latent distributions;
• Unlocks 1.8x to 2.4x end-to-end throughput acceleration when paired with speculative decoding engines;

1. Unencumbered Apache 2.0 Licensing:

• All weights, tokenizers, and speculative artifacts are published under the permissive Apache 2.0 license, permitting commercialization and domain-specific fine-tuning.

Benchmarks and Measured Results

Based on Krafton AI's technical disclosures and verified playground benchmarks:

• Parity with Meta's Muse Glimmer:
• Matches Meta's recently released Muse Glimmer across agentic multi-tool execution and complex software refactoring benchmarks;
• Consumer Hardware Footprint:
• Thanks to 4B sparse activation, quantized weights run fluidly inside 16GB VRAM GPUs, cutting memory bandwidth pressure by over 60% relative to dense 30B alternatives;
• Streaming Turnaround Velocity:
• In speculative generation mode, Raon-1 mini sustains interactive streaming speeds of 80 to 120 tokens/sec on local consumer GPUs.

Getting Started for Developers

• Interactive Playground: Explore live demonstrations at https://raon-mini.krafton.ai/en;
• Official Announcement: View verified thread via @Krafton_AI status https://x.com/Krafton_AI/status/2112772568776339709;
• Deployment Strategy: Developers building game NPC intelligences or air-gapped terminal assistants can serve Raon-1 mini via Ollama or vLLM on standard gaming laptops and workstations without cloud connectivity.