[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fuau5of06o9pj":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":24,"bodyZh":31,"bodyEn":32,"sourceName":33,"sourceUrl":34,"updatedAt":35},"krafton-raon-1-mini-2026-10-22","NEWS","2026-10-22T14:00:00.000Z","KRAFTON Raon-1-mini 预览版：28.8B-A4.3B 混合注意力 MoE，NVFP4 单卡 5090 可跑","KRAFTON Raon-1-mini preview: a 28.8B-A4.3B hybrid-attention MoE that fits a single RTX 5090 at NVFP4","10 月 22 日 KRAFTON AI 公开 Raon-1-mini 预览版（Apache-2.0，申请制自动批准）——从零训练的韩语强项 MoE：28.8B 总参 \u002F 4.3B 激活，25 层里 18 层 Raon Delta Attention（线性注意力）+ 7 层 Gated MLA（KV 约 8KiB\u002Ftoken），262K 上下文。官方 NVFP4 单卡 RTX 5090 可推理，DFlash2 草稿模型 B300 实测最高 2.18 倍吞吐。官方 GGUF Q4_K_M 19.7GB。站内新增 24 条实测。","KRAFTON AI published the Raon-1-mini preview on Oct 22 (Apache-2.0, request-to-access with automatic approval) — a from-scratch MoE strong at Korean: 28.8B total \u002F 4.3B active, with 18 of its 25 layers using Raon Delta Attention (linear attention) and 7 using Gated MLA (KV of ~8KiB per token), 262K context. Official NVFP4 runs on a single RTX 5090; the DFlash2 draft model delivers up to 2.18x throughput on a B300. Official GGUF Q4_K_M is 19.7GB. 24 new on-site runs.",[12,15,18,21],{"id":13,"name":14},5,"Qwen3.8-27B",{"id":16,"name":17},47,"Qwen3.6-35B-A3B",{"id":19,"name":20},93,"Raon-1-mini-preview",{"id":22,"name":23},95,"elephant-1",[25,28],{"id":26,"name":27},13,"NVIDIA RTX 5090",{"id":29,"name":30},110,"NVIDIA RTX Spark","# KRAFTON Raon-1-mini 预览版：28.8B-A4.3B 混合注意力 MoE，NVFP4 单卡 5090 可跑\n\n10 月 22 日，PUBG 母公司 KRAFTON 的 AI 子公司公开了 Raon 开放模型家族的首款——**Raon-1-mini 预览版**（Apache-2.0）。这是少数「从零训练」的非中美系开放权重模型：28.8B 总参 \u002F 4.3B 激活（官方记号 30B-A4B），韩语会话与 agentic 任务是主攻方向。韩国媒体（Chosunbiz \u002F MoneyToday \u002F 首尔经济 \u002F AsiaE）连续两天跟进报道。\n\n先说清状态：**这是 preview，不是最终版**——官方明确「规格与结果在正式版前可能变化」；权重仓库申请制但自动批准。\n\n## 混合注意力：18 层线性 + 7 层 MLA\n\n25 层里只有 7 层保留 KV cache：\n\n| 组件 | 配置 |\n|---|---|\n| Raon Delta Attention | 18 层，固定大小状态（把 Kimi Delta Attention 的 erase \u002F write 门解耦） |\n| Gated MLA | 7 层，KV 仅约 8KiB\u002Ftoken（BF16） |\n| 排布 | 3:1 混排，40 头，头维 128（RDA）\u002F 192QK-128V（MLA） |\n| MoE | 128 路由专家选 6 + 1 共享 = 每 token 4.3B |\n| 上下文 | 262,144 token |\n\n线性注意力层的状态不随上下文增长，MLA 层把 KV 压到 8KiB\u002Ftoken——两者叠加后 262K 上下文的显存账几乎只剩权重本身。这是目前开放权重里最激进的混合注意力方案之一。\n\n## NVFP4：单卡 5090 可跑\n\n官方量化提供 **MXFP8 与 NVFP4** 两种 checkpoint，NVFP4 让整模型进入单张 RTX 5090（32GB）。另附 **DFlash2 草稿模型**（官方 B300 实测，vLLM 0.28.0 \u002F block 8）：输出吞吐最高 **2.18 倍**。tool calling 原生支持，Hermes Agent、Pi、OpenCode 已接入。\n\n## 官方 GGUF（KRAFTON 仓库，10-22）\n\n| 量化 | 文件大小 |\n|---|---|\n| IQ4_XS | 16.1 GB |\n| Q4_K_M | 19.7 GB |\n| Q5_K_M | 22.2 GB |\n| Q6_K | 25.8 GB |\n| Q8_0 | 30.7 GB |\n| DFlash2 草稿（Q8_0） | 0.7 GB |\n\nQ8_0 30.7GB 贴着 5090 的 32GB 上限，Q4_K_M 19.7GB 是 24GB 卡的现实档位。同样提醒：站内显存口径含 padding 余量，保守公式下 4.3B 激活的 28.8B 总参要更大的卡才判 fit——物理 GGUF 与站内 fit 判定是两本账，各看各的。\n\n## 站内收录\n\n**Raon-1-mini-preview**（id 93）已入库，新增 **24 条 L2 实测**，全池均值 136.1 tok\u002Fs：\n\n| 硬件 | 量化 | 实测 |\n|---|---|---|\n| RTX 5090 | Q2_K | 351.2 tok\u002Fs |\n| H100 | Q6_K | 343.5 tok\u002Fs |\n| RTX PRO 6000 96GB | Q4_K_M | 227.9 tok\u002Fs |\n| M2 Ultra 80GB | Q4_K_M | 110.5 tok\u002Fs |\n| RTX Spark（116 GiB Linux 池） | Q4_K_M | 45-49 tok\u002Fs |\n\nA4.3B 激活介于 A3B 与稠密 27B 之间，实测曲线与站内 Qwen3.6-35B-A3B 的经验一致——MoE 解码速度基本由激活参数 × 卡带宽决定，总参只决定装不装得下。","# KRAFTON Raon-1-mini preview: a 28.8B-A4.3B hybrid-attention MoE that fits a single RTX 5090 at NVFP4\n\nOn Oct 22, KRAFTON AI — the AI arm of the PUBG parent company — published the first model in the open Raon family, the **Raon-1-mini preview** (Apache-2.0). It is one of very few non-US\u002FChina open-weight models trained from scratch: 28.8B total \u002F 4.3B active (official designation 30B-A4B), aimed squarely at Korean conversation and agentic tasks. Korean outlets (Chosunbiz, MoneyToday, Seoul Economic, AsiaE) covered it wall-to-wall for two days.\n\nStatus first: **this is a preview, not the final model** — the official line is that specs and results may change before release. Weight repos are request-to-access with automatic approval.\n\n## Hybrid attention: 18 linear layers + 7 MLA\n\nOnly 7 of the 25 layers keep a KV cache:\n\n| Component | Configuration |\n|---|---|\n| Raon Delta Attention | 18 layers, fixed-size state (decouples the erase\u002Fwrite gates of Kimi Delta Attention) |\n| Gated MLA | 7 layers, KV of only ~8KiB\u002Ftoken (BF16) |\n| Layout | 3:1 interleaved, 40 heads, head dim 128 (RDA) \u002F 192QK-128V (MLA) |\n| MoE | 6 of 128 routed experts + 1 shared = 4.3B per token |\n| Context | 262,144 tokens |\n\nLinear-attention state does not grow with context and MLA squeezes KV to 8KiB\u002Ftoken — together the 262K-context VRAM bill is almost nothing but weights. This is among the most aggressive hybrid-attention designs in open weights today.\n\n## NVFP4: a single RTX 5090\n\nOfficial quantizations ship as **MXFP8 and NVFP4** checkpoints, and NVFP4 puts the whole model on a single RTX 5090 (32GB). A **DFlash2 draft model** also ships (official B300 numbers, vLLM 0.28.0 \u002F block 8): up to **2.18x** output throughput. Tool calling is native, with Hermes Agent, Pi and OpenCode already integrated.\n\n## Official GGUFs (KRAFTON repo, Oct 22)\n\n| Quant | File size |\n|---|---|\n| IQ4_XS | 16.1 GB |\n| Q4_K_M | 19.7 GB |\n| Q5_K_M | 22.2 GB |\n| Q6_K | 25.8 GB |\n| Q8_0 | 30.7 GB |\n| DFlash2 draft (Q8_0) | 0.7 GB |\n\nQ8_0 at 30.7GB sits right against the 5090's 32GB; Q4_K_M at 19.7GB is the realistic tier for 24GB cards. Same reminder as above: the site's padded VRAM formula judges fit more conservatively — physical GGUF size and the on-site fit verdict are two different ledgers.\n\n## On-site coverage\n\n**Raon-1-mini-preview** (id 93) is cataloged with **24 new L2 records**, averaging 136.1 tok\u002Fs across the pool:\n\n| Hardware | Quant | Measured |\n|---|---|---|\n| RTX 5090 | Q2_K | 351.2 tok\u002Fs |\n| H100 | Q6_K | 343.5 tok\u002Fs |\n| RTX PRO 6000 96GB | Q4_K_M | 227.9 tok\u002Fs |\n| M2 Ultra 80GB | Q4_K_M | 110.5 tok\u002Fs |\n| RTX Spark (116 GiB Linux pool) | Q4_K_M | 45-49 tok\u002Fs |\n\nAt 4.3B active it sits between A3B and dense 27B, and the measured curve matches what we already know from Qwen3.6-35B-A3B on-site: MoE decode speed is essentially active params times card bandwidth — total params only decide whether it fits.","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fkrafton-raon-1-mini-2026-10-22","2026-10-22T15:23:50.385Z"]