[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-krafton-raon-1-mini-preview-open-weights":3,"topics-all":38,"news-related-03d33c23-e56a-41eb-8732-b0712059662f":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"03d33c23-e56a-41eb-8732-b0712059662f","PUBG 母公司开源 MoE:激活仅 4.3B 单卡可跑","韩国游戏公司 KRAFTON 发布首个开放权重模型 Raon-1 mini preview:总参 28.8B、每 token 激活 4.3B 的 MoE,原生 262K 上下文,Apache 2.0 许可。NVFP4 量化后单张 RTX 5090 可推理,官方称韩语与 agent 任务表现突出。","PUBG 的母公司 KRAFTON 在 Hugging Face 上放出了第一个开放权重语言模型 Raon-1 mini preview——不是拿现成模型微调,而是官方口径里「从零设计、从零训练」的 30B-A4B 稀疏 MoE,Apache 2.0 许可,权重直接可下。游戏大厂下场做基座模型,这件事本身比跑分更值得看一眼。\n\n## 一个把当红组件缝满的 30B\n\n规格表先摆出来:总参 28.8B,每 token 激活 4.3B——128 个路由专家里每 token 只走 6 个,外加 1 个共享专家;25 层,hidden size 2560。注意力是最有意思的部分:18 层线性注意力(RDA)+ 7 层 Gated MLA 的 3:1 混合布局,只有那 7 层 MLA 保留 KV cache,官方给的记忆开销约每 token 8 KiB(BF16),这也是它原生 262,144 token 上下文的底气。\n\n两个自研命名的组件都站在别人的肩膀上:Raon Delta Attention 官方描述是「解耦 Kimi Delta Attention 的 erase 和 write 门」,以改进长上下文表现;Raon-mHC 则是把 DeepSeek 的流形约束超连接(mHC)做降本改造,官方称这能加速训练。加上 NoPE 无位置编码与四流残差,这个 30B 几乎把 2026 年当红的架构组件逐个魔改了一遍——小团队做架构实验的标准姿势:不发明轮子,把轮子改到认不出来。\n\n## 部署画像:一张 RTX 5090 装得下\n\n量化给得很全:官方发布 MXFP8 和 NVFP4 两种格式,NVFP4 版可在单张 RTX 5090(32GB 显存)上推理;聚合页 theopenweights 按权重元数据估算,FP8 版本约 31.9GB,4-bit 下 24GB 显卡(RTX 3090\u002F4090)也能跑。另外还配了一个 DFlash2 draft 模型做投机解码加速。对延迟敏感的端侧和游戏内应用,4.3B 激活的推理成本画像相当务实。\n\n## 官方自报的部分,先打个折\n\n模型卡的 Highlights 全是官方口径:内置 thinking,数学与科学基准「在对比模型中领先」;指令遵循与搜索类基准领先;韩、英、日、中、西、法、泰七语种训练,官方称韩语表现最强;agent 工具调用支持 Hermes Agent、Pi 和 OpenCode。但注意两点:README 明确写了这是 preview,「specs and results may change」;模型卡上的跑分属官方自报,最终版之前一切以实际复现为准。当前 HF 下载量 249、GitHub 3 stars,社区热度还没起来。\n\n## 所以呢\n\ntheopenweights 的评语很准:这不是即插即用的成品,而是「意图信号」。有意思的观察点在后面:一家以交互娱乐见长的公司,从零训出一个 MoE 之后会把它优化向哪个方向,素材没有给答案,final 版会说话。当做游戏的公司开始从零训 MoE,开放权重这张桌子的玩家名单,比多数人想象的要长。\n\n参考:[KRAFTON\u002FRaon-1-mini-preview 模型卡](https:\u002F\u002Fhuggingface.co\u002FKRAFTON\u002FRaon-1-mini-preview) · [GitHub 仓库](https:\u002F\u002Fgithub.com\u002Fkrafton-ai\u002FRaon-1-mini-preview)","https:\u002F\u002Fhuggingface.co\u002FKRAFTON\u002FRaon-1-mini-preview","24d5c6c5-6573-4180-a1fd-f1459842d1af",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"994fea84-fe56-4205-8fe9-a143c87f4ef5","en","KRAFTON of PUBG Open-Sources 30B-A4B MoE Raon-1 mini","PUBG studio KRAFTON open-sources Raon-1 mini: a 28.8B MoE with 4.3B active, 262K context, Apache 2.0 license. NVFP4 runs on a single RTX 5090.","KRAFTON, the South Korean game company behind PUBG, has quietly put its first open-weight language model on Hugging Face: Raon-1 mini preview. This is not a fine-tune of someone else's checkpoint — the model card describes it as \"designed and trained from scratch by KRAFTON AI\", a 30B-A4B sparse MoE under the Apache 2.0 license. A major game studio entering the base-model arena is, by itself, more interesting than any benchmark line.\n\n## A 30B stuffed with 2026's favorite components\n\nThe spec sheet: 28.8B total parameters, 4.3B activated per token — only 6 of 128 routed experts fire per token, plus 1 shared expert; 25 layers, hidden size 2,560. The attention design is the most interesting part: a 3:1 hybrid of 18 linear-attention layers (RDA) and 7 Gated MLA layers. Only those 7 MLA layers keep a KV cache, at roughly 8 KiB per token in BF16 — which is where the native 262,144-token context comes from.\n\nBoth custom-named components stand on other people's shoulders. Raon Delta Attention, per the official description, decouples the erase and write gates of Kimi Delta Attention to improve long-context performance. Raon-mHC is a cost-reduced take on DeepSeek's manifold-constrained hyper-connections, which the team says speeds up training. Add NoPE (no positional encoding) and a 4-stream residual, and this 30B remixes nearly every fashionable architecture trick of 2026 — the standard playbook for a small team: don't invent the wheel, just modify it beyond recognition.\n\n## The deployment story: fits on one RTX 5090\n\nQuantization options are generous: official MXFP8 and NVFP4 builds, with the NVFP4 version running inference on a single NVIDIA RTX 5090 (32 GB VRAM). Aggregator theopenweights, reading the weights metadata, puts the FP8 build at about 31.9 GB and estimates the 4-bit build fits on 24 GB cards (RTX 3090\u002F4090). A DFlash2 draft model ships alongside for accelerated speculative decoding. For latency-sensitive on-device and in-game applications, a 4.3B-active inference profile is a fairly pragmatic choice.\n\n## The officially claimed parts: apply a discount\n\nThe highlights on the model card are all vendor claims: built-in thinking with \"leading results\" on the math and science benchmarks shown; leads on instruction-following and search benchmarks; trained on Korean, English, Japanese, Chinese, Spanish, French and Thai, with the strongest Korean performance; tool calling for agentic workflows with support for Hermes Agent, Pi and OpenCode. Two caveats: the README states plainly that this is a preview and \"specs and results may change\", and the benchmark tables are self-reported — treat everything as provisional until the final release and independent reproduction. Current traction is modest: 249 downloads on Hugging Face, 3 stars on GitHub.\n\n## So what\n\ntheopenweights called it right: this is \"more of a signal of intent than a drop-in replacement\". The interesting question comes next — a company rooted in interactive entertainment has trained a MoE from scratch; which direction will it optimize toward? The published record gives no answer yet; the final release will. When the company behind PUBG starts training its own MoE, the guest list at the open-weights table turns out to be longer than most people assumed.\n\nRefs: [KRAFTON\u002FRaon-1-mini-preview model card](https:\u002F\u002Fhuggingface.co\u002FKRAFTON\u002FRaon-1-mini-preview) · [GitHub repo](https:\u002F\u002Fgithub.com\u002Fkrafton-ai\u002FRaon-1-mini-preview)","krafton-raon-1-mini-preview-open-weights","2026-10-23T13:08:33Z","2026-10-23T13:08:37.110853Z","2026-10-23T13:08:37.110863Z",true,"agent",11,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"bbfcbee4-8d92-4078-bb1e-eaf4d1c1e146","Beam 权重落地:501B MoE 三精度齐发","beam-1-0-weights-open-agent-stack","2026-10-18T21:10:09+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"cee60a45-f212-4a78-83dd-0226c07b3460","腾讯开源记忆写入模型 Hy-Memory:30B MoE 只激活 3B,四框架平均分超 Opus 5","tencent-hy-memory-30b-memory-writer","2026-10-17T15:14:21+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"ac666c4e-4756-43d8-840a-4eacbf88f804","Meta 开源 816B MoE 模型 Muse Spark Open,跑在两台 Mac Studio 上","meta-muse-spark-open-816b-mac-studio","2026-10-17T03:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"1b6ec126-cba1-47da-870b-da19cc523763","Meta 开源 816B MoE 模型 Muse Spark Open:1M 上下文,两台 Mac Studio 跑得动","meta-muse-spark-open-816b-moe-1m-mac-studio","2026-10-15T11:00:00+00:00"]