Raon-1 mini preview

Raon-1 mini preview is a 28.78B mixture-of-experts (MoE) language model built from scratch by KRAFTON AI. Operating with only 4.3B active parameters per token, this architecture places the model firmly in the 30B-A4B efficiency tier. It is designed to bridge the gap between heavy frontier reasoning models and lean, locally deployable architectures, focusing heavily on native Korean comprehension, fluent English generation, multi-step agentic execution, and long-form document processing.

Released under a permissive Apache 2.0 license, Raon-1 mini preview is an open-weight release that gives developers full commercial and private deployment rights. While many open models treat Korean as a secondary translation target, KRAFTON trained this base architecture with Korean at its foundation alongside strong multilingual capabilities. Practitioners looking for a capable local AI model 28.78B parameters 2026 deployment will find that this preview release punches significantly above its weight class in instruction following, autonomous search, and coding tasks.

Because this is explicitly designated as a preview release, developers should expect iterative improvements to weights, prompt templates, and system prompts prior to the final release. However, its architectural foundation already presents a compelling case for local deployment on modern workstation hardware.

The core strength of Raon-1 mini preview lies in its hybrid design, combining sparse routing with an asymmetric attention layout. Rather than relying entirely on standard transformer self-attention, KRAFTON employs a 25-layer configuration made up of 18 linear attention layers and 7 full attention layers.

• Hybrid Attention Pipeline: 18 layers use Raon Delta Attention (RDA), a mechanism that decouples the erase and write gates found in traditional linear attention schemes (such as KDA). This allows the linear layers to maintain a fixed-size internal state without ballooning computational costs over massive context windows.
• Compact Key-Value Cache: Only the 7 full attention layers use Multi-Head Latent Attention (MLA) and require an active Key-Value (KV) cache. At BF16 precision, this requires approximately 8 KiB per token. This keeps memory overhead exceptionally small even when scaling conversations out toward the native 262,144 token limit.
• Sparse Routing (6 of 128 Experts): Raon-1 mini preview MoE efficiency comes from routing each token through 6 active experts out of 128 total routed options. While your hardware must hold the entire 28.78B parameter weight footprint in memory, execution compute is restricted to just 4.3B parameters per forward pass.
• Raon-mHC Connections: The network integrates manifold-constrained hyper-connections to stabilize gradient propagation and accelerate training throughput without adding latency during inference.
• Speculative Decoding Support: KRAFTON provides a native DFlash2 draft model alongside the base weights. This allows local inference engines to draft candidate tokens using a fast auxiliary model, speeding up sequential token generation.

Raon-1 mini preview includes a built-in thinking mode that generates explicit chain-of-thought traces before returning final answers. This makes it particularly effective for technical reasoning, formal logic, and complex step-by-step problem solving.

The model natively supports multi-turn tool calling and schema adherence. It integrates smoothly with agent environments like Hermes Agent, Pi, and OpenCode. When executing agent workflows, Raon-1 mini preview reads structured API outputs, handles schema validation, and chains tool responses back into intermediate reasoning steps without hallucinating non-existent function arguments.

The model demonstrates high fidelity on agentic coding benchmarks such as SWE-Verified, where models must clone a repository, locate a bug, modify code across files, and run tests. Thanks to its native 262,144 token context window, you can feed entire repository modules, build logs, and API definitions into the prompt without truncating context. It writes idiomatic Python, TypeScript, C++, and Go, handling everything from algorithm generation to debugging unit tests.

While optimized for deep Korean syntax and cultural context, the training corpus also includes English, Japanese, Chinese, Spanish, French, and Thai. On academic science benchmarks like GPQA Diamond and factuality tests like AA-Omniscience, the model delivers strong reasoning benchmark results among sub-35B MoE architectures. It also excels on multi-document synthesis tests (such as BrowseComp Plus), accurately gathering facts across scattered web documents while declining to answer when data is missing.

To run Raon-1 mini preview locally, your hardware planning must account for two distinct factors: the full 28.78B weight footprint required in memory, and the KV cache sizing needed for long-context runs.

Because MoE models must store all experts in memory even though only 4.3B are active at any given step, VRAM capacity is your main deployment bottleneck:

• NVFP4 (4-bit native / quant): ~18 GB to 22 GB VRAM. Fits entirely within a 24 GB card (such as an RTX 3090 or RTX 4090) for standard context, or a 32 GB RTX 5090 with full 262K context headroom.
• MXFP8 / FP8 (8-bit): ~32 GB to 36 GB VRAM. Requires dual 24 GB GPUs (48 GB total VRAM) or an Apple Silicon Mac with 48 GB+ unified memory.
• BF16 / FP16 (Unquantized): ~60 GB to 65 GB VRAM. Requires dual 48 GB enterprise cards (like RTX 6000 Ada or A6000) or an Apple Silicon system with 64 GB to 128 GB unified memory.

When choosing the best GPU for Raon-1 mini preview, your target context length determines the optimal setup:

• Single NVIDIA RTX 5090 (32 GB): The premier single-GPU target. With NVFP4 quantization, the entire model fits onto the RTX 5090 while leaving ample memory for the 262K context window.
• Dual NVIDIA RTX 3090 / RTX 4090 (2x 24 GB = 48 GB): Ideal for running FP8 or high-precision 6-bit quantizations across standard PCIe lanes via pipeline or tensor parallelism in vLLM.
• Apple Silicon (Mac Studio / MacBook Pro M3/M4 Max with 64 GB or 128 GB): The easiest platform for unquantized or 8-bit local testing without dealing with multi-GPU bifurcation.

For everyday local workloads, the best quantization for Raon-1 mini preview is either the officially supplied NVFP4 format on Blackwell hardware or a balanced Q4_K_M GGUF format on existing systems. Because the KV cache only grows across 7 of the 25 layers, long-context prompts produce a minimal memory penalty compared to standard dense transformers.

Because the active parameter load is just 4.3B tokens, Raon-1 mini preview tokens per second generation speeds match those of small dense models:

• RTX 5090 (NVFP4): Expected generation speeds run between 65 and 90 tokens per second for short outputs, settling around 45 to 60 tokens per second during deep speculative decoding loops.
• Dual RTX 4090 (vLLM / FP8): Expect roughly 50 to 70 tokens per second depending on prompt-processing lengths.
• Apple Silicon M4 Max (GGUF Q4_K_M via llama.cpp): Generates roughly 40 to 55 tokens per second, making it an exceptionally fast local assistant.

For single-line deployment through Ollama once community quantizations are imported:

1# Pull and run the model via Ollama (once GGUF weights are converted)
2ollama run raon-1-mini-preview

For production agent workflows requiring OpenAI-compatible endpoints with tool calling and high concurrency, deploy using vLLM on Linux:

1vllm serve KRAFTON/Raon-1-mini-preview \
2  --tensor-parallel-size 1 \
3  --quantization nvfp4 \
4  --max-model-len 32768 \
5  --enable-auto-tool-choice \
6  --tool-call-parser hermes

If you are running multi-GPU configurations on older cards, increase --tensor-parallel-size to match your GPU count and switch quantization to standard FP8 or AWQ.

Evaluating Raon-1 mini preview vs competitor architectures highlights where KRAFTON's engineering choices pay off:

• Raon-1 mini preview vs Qwen2.5-32B: Qwen2.5-32B is a dense model, meaning every single token processes across all 32 billion parameters. This demands massive compute throughput, limiting generation speeds on consumer setups to roughly 25 to 35 tokens per second. Raon-1 mini preview demands similar total VRAM to hold weights, but generates at nearly double the speed because it only calculates 4.3B active parameters per forward pass. Furthermore, Raon provides a massive native 262K context window versus Qwen's native 128K window, with significantly stronger default Korean phrasing.
• Raon-1 mini preview vs Gemma-2-27B: Gemma 2 27B offers exceptional English reasoning but struggles with Korean nuances, lacks built-in dynamic tool-calling scaffolding, and is capped at an 8,192 token context window. Raon-1 mini preview easily outperforms Gemma 2 on agentic workflows, long-document digestion, and multilingual reasoning tasks, while keeping KV cache growth flat across deep conversations.

Choose Raon-1 mini preview if your workflow involves Korean linguistic processing, autonomous agents with external tool calling, or deep document analysis requiring extended context lengths on single-workstation GPUs.