# Raon-1 mini preview ![Organization](https://img.shields.io/badge/Org-KRAFTON-9333ea) ![Context](https://img.shields.io/badge/Context-262K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Apache%202.0-2da44e) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) ![Released](https://img.shields.io/badge/Released-Oct%202026-555) > **Quick answer:** Raon-1 mini preview is KRAFTON AI's first open language model — a 28.8B-parameter MoE (4.3B active) with a hybrid linear/full-attention architecture (Raon Delta Attention + Gated MLA), Apache 2.0 licensed, with a 262,144-token context window. It targets Korean conversation, reasoning, and agentic tool use: 84.3% GPQA Diamond, 93.3% AIME 2026, and 92.9% on KRAFTON's internal Korean chatbot eval. ## At a Glance **Where Raon-1 mini preview leads** - Instruction following: 84.2% IFBench — ahead of all MoE/dense comparison models shown - Korean conversation: 92.9% on KRAFTON's internal Korean chatbot benchmark - Reasoning: 84.3% GPQA Diamond, 93.3% AIME 2026 — best among compared MoE models (Qwen3.6-35B-A3B, Gemma 4 26B-A4B) - Efficient hybrid attention: only 7 of 25 layers keep a KV cache (~8 KiB/token in BF16), enabling long context at low memory cost **Where it lags** - Terminal-Bench Lite: 47.0% — well behind both MoE and dense comparison models - Knowledge: 76.5% MMLU Pro — behind Qwen3.6-35B-A3B (85.9%) and dense comparison models - Preview status: KRAFTON states specs and results may change before the final release **Best for:** Korean-language conversational and agentic workloads needing strong instruction-following at a compact MoE active-parameter cost. ## What Raon-1 mini preview Is Raon-1 mini preview is KRAFTON AI's first entry in its open "Raon" model family, trained from scratch rather than fine-tuned from an existing base. It uses a hybrid Mixture-of-Experts architecture with 25 layers: 18 Raon Delta Attention (RDA) layers — a variant of Kimi Delta Attention that decouples erase and write gates for better long-context retention — and 7 Gated MLA (full-attention) layers that are the only ones carrying a KV cache. Of 128 routed experts, 6 plus 1 shared expert activate per token (4.3B of 28.8B total parameters). The model is explicitly multilingual-but-Korean-first: trained on Korean, English, Japanese, Chinese, Spanish, French, and Thai, with KRAFTON reporting particularly strong Korean colloquial performance on an internal chatbot benchmark alongside solid English results. It ships with built-in "thinking" (reasoning) mode that KRAFTON recommends keeping enabled, tool-calling support for agentic frameworks (Hermes Agent, Pi, OpenCode), and a companion DFlash2 draft model for speculative decoding (up to ~2.18x throughput on a single NVIDIA B300). As a preview release, KRAFTON cautions that specs and benchmark results may still change before a final version ships. ## Specifications | Field | Value | |---|---| | Organization | KRAFTON AI | | Parameters | 28.8B total; 4.3B active per token (MoE, 128 routed experts, 6 routed + 1 shared active) | | Context window | 262,144 tokens | | Architecture | Hybrid MoE — 18 Raon Delta Attention (linear, no KV cache) + 7 Gated MLA (full-attention) layers, NoPE, Raon-mHC residual connections | | License | Apache 2.0 | | Release date | October 2026 (preview) | | Modality | Text | ## Pricing Open weights — free to download from Hugging Face; self-hosting cost only (gated repository, automatic approval). Quantized MXFP8 and NVFP4 checkpoints are available; NVFP4 runs on a single NVIDIA RTX 5090 (32GB VRAM). ## Context Window Raon-1 mini preview natively supports **262,144 tokens** of context. Only the 7 Gated MLA layers maintain a KV cache (about 8 KiB per token in BF16); the 18 Raon Delta Attention layers use a fixed-size state instead, which KRAFTON reports improves long-context performance over standard Kimi Delta Attention while keeping memory costs low at long context lengths. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [MMLU Pro](/benchmarks/wang-et-al/mmlu-pro) | 76.5 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) | 84.3 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [AIME 2026](/benchmarks/mathematical-association-of-america/aime-2026) | 93.3 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [HMMT Feb 2026](/benchmarks/hmmt/hmmt-2026) | 84.1 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [IFBench](/benchmarks/ifbench/ifbench) | 84.2 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [MultiChallenge](/benchmarks/multichallenge/multichallenge) | 59.0 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [LiveCodeBench v6](/benchmarks/livecodebench/livecodebench-v6) | 81.6 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [SciCode](/benchmarks/scicode/scicode) | 38.7 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [SWE-bench Verified](/benchmarks/princeton-nlp-openai/swe-bench-verified) | 70.8 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [BFCL v4](/benchmarks/bfcl/bfcl-v4) | 59.8 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [AA-LCR](/benchmarks/artificial-analysis/aa-lcr) | 67.3 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [AA-Omniscience](/benchmarks/artificial-analysis/aa-omniscience) | 3.5 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | | [AutomationBench](/benchmarks/zapier/automationbench) | 29.7 | [Raon-1 mini preview model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview) | 2026-10 | *Scores above are reported by KRAFTON and shown for context. They are not Benchgen measurements. Figures depend on harness, tools, and effort settings — see the source for methodology. SWE-bench Verified score was evaluated by KRAFTON using the mini-SWE-agent harness.* **Excluded vendor scores (not mapped to any local benchmark entity):** Terminal-Bench Lite (47.0%) — distinct from locally-tracked Terminal-Bench 3/4/Hard/Science variants, no confirmed shared methodology; GDPVal (66.0, judge score) — KRAFTON's table explicitly notes this is a judge score, not an Elo rating, so it does not map onto GDPVal-AA v2's Elo scale; BrowseComp Plus (55.7%) — no matching local entity for this "Plus" variant distinct from the base BrowseComp; GDP.pdf (5.8, text-only OCR'd version) — ambiguous metric/scale versus GDPVal-AA v2, same exclusion precedent applied to other vendors' GDP.pdf scores this cycle. Each excluded per the scope-assumption rule rather than guessed. ## Raon-1 mini preview vs Alternatives | Model | Context | GPQA Diamond | AIME 2026 | License | |---|---|---|---|---| | **Raon-1 mini preview** | 262K | 84.3 | 93.3 | Apache 2.0 | | [Qwen3.6-35B-A3B](/models/alibaba/qwen3-6-35b-a3b) | — | 83.2 | 92.9 | Apache 2.0 | | [Gemma 4 26B-A4B](/models/google/gemma-4-26b-a4b) | — | 80.7 | 87.1 | Gemma Terms of Use | Among compared MoE models, Raon-1 mini preview leads on both knowledge-reasoning benchmarks shown here while using a smaller active-parameter budget (4.3B) than Qwen3.6-35B-A3B — KRAFTON's hybrid attention design appears to buy reasoning quality without a larger active-parameter cost, though it trails on raw knowledge recall (MMLU Pro) and terminal-agent tasks. ## How Raon-1 mini preview Performs on Real Agent Tasks KRAFTON reports agentic signals across SWE-bench Verified (code fixing, mini-SWE-agent harness), BFCL v4 (tool calling), and AutomationBench (business workflow automation) — a reasonably broad agentic footprint for a first-generation model. The large gap between its reasoning/instruction-following strength and its weaker Terminal-Bench Lite and AutomationBench scores suggests the model's agentic tool-use reliability lags its raw reasoning capability; builders evaluating it for autonomous coding or workflow-automation tasks should validate against their own task distribution rather than extrapolating from the knowledge/reasoning numbers alone. ## Use Raon-1 mini preview via API ```python # Raon-1 mini preview ships open weights; serve locally with vLLM (beta, KRAFTON fork) from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY") response = client.chat.completions.create( model="Raon-1-mini-preview", messages=[{"role": "user", "content": "Summarize this contract clause..."}], extra_body={"chat_template_kwargs": {"enable_thinking": True}}, ) print(response.choices[0].message.content) ``` ## Related Models & Benchmarks - Compare: [Qwen3.6-35B-A3B](/models/alibaba/qwen3-6-35b-a3b) · [Gemma 4 26B-A4B](/models/google/gemma-4-26b-a4b) - Benchmarks: [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) · [AIME 2026](/benchmarks/mathematical-association-of-america/aime-2026) · [SWE-bench Verified](/benchmarks/princeton-nlp-openai/swe-bench-verified) ## Frequently Asked Questions

What is Raon-1 mini preview?

KRAFTON AI's first open language model — a 28.8B-parameter (4.3B active) hybrid-attention Mixture-of-Experts model trained from scratch, built for Korean conversation, reasoning, and agentic tool use. It is a preview release; specs may change before the final version.

What is Raon-1 mini preview's context window?

262,144 tokens — enough for large documents or extended multi-turn agent conversations in a single request.

How much does Raon-1 mini preview cost?

Open weights — free to download from Hugging Face. Self-hosting cost only; NVFP4 quantization runs on a single NVIDIA RTX 5090 (32GB).

Is Raon-1 mini preview open source?

Yes — released under the Apache 2.0 license.

What is Raon-1 mini preview's knowledge cutoff?

Not disclosed in the public model card.

--- *Specs and scores sourced from KRAFTON AI's official Raon-1 mini preview model card (October 2026). Last updated 2026-10-25.*

Raon-1 mini preview — Training jobs

Training jobs