{"id":7682,"date":"2025-11-03T08:52:00","date_gmt":"2025-11-02T23:52:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7682"},"modified":"2025-11-02T13:08:25","modified_gmt":"2025-11-02T04:08:25","slug":"minimax-m2-kimi-linear-ling-v2-ouro-emu3-5-gpt-oss-safeguard","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7682","title":{"rendered":"MiniMax M2, Kimi-Linear,  Ling-V2, Ouro, Emu3.5, gpt-oss-safeguard"},"content":{"rendered":"\n<p>\u5148\u9031\u306f\u516c\u958b\u30e2\u30c7\u30eb\u306e\u8a71\u984c\u304c\u591a\u304f\u3001\u305d\u306e\u4e2d\u3067\u3082<a href=\"https:\/\/huggingface.co\/MiniMaxAI\/MiniMax-M2\">MiniMax-M2 <\/a>\u3068<a href=\"https:\/\/github.com\/MoonshotAI\/Kimi-Linear\">Kimi-Linear<\/a>\u306f\u8981\u6ce8\u76ee\u3002\u7279\u306b\u5f8c\u8005\u306f\u52b9\u7387\u6027\u3082\u9ad8\u3044\u3002\u5148\u9031\u306eRing\u3068\u3084\u3084\u3053\u3057\u3044\u304c\u3001<a href=\"https:\/\/github.com\/inclusionAI\/Ling-V2\">Ling-V2<\/a>\u3082\u5f37\u529b\u306a\u30e2\u30c7\u30eb\u3067\u3042\u308b\uff08This report focuses on three reflex-grade non-thinking (instruct) models in the Ling 2.0 family\u2014Ling-mini-2.0, Ling-flash-2.0, and Ling-1T. These models emphasize general reasoning and instruction-following capability, while the Ring series (Ling-Team, 2025), built upon the same Ling 2.0 base, extends toward deep thinking models.\u3068\u306e\u3053\u3068\uff09\u3002\u307e\u305f\u3001\u5c0f\u578b\u30e2\u30c7\u30eb\u3067\u3042\u308b<a href=\"https:\/\/huggingface.co\/ByteDance\/Ouro-2.6B\">Ouro-2.6B <\/a>\u3001<a href=\"https:\/\/huggingface.co\/ByteDance\/Ouro-2.6B-Thinking\">Ouro-2.6B-Thinking<\/a>\u3082\u8208\u5473\u6df1\u304b\u3063\u305f\u3002<\/p>\n\n\n\n<p>\u4e0a\u8a18\u3068\u306f\u7570\u306a\u308b\u304c\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30eb\u306a<a href=\"https:\/\/github.com\/baaivision\/Emu3.5\">Emu3.5<\/a>\u3001\u5206\u985e\u30bf\u30b9\u30af\uff08safety classification tasks\uff09\u7528\u306e<a href=\"https:\/\/huggingface.co\/collections\/openai\/gpt-oss-safeguard\">gpt-oss-safeguard<\/a>\u306a\u3069\u5f37\u529b\u306a\u30e2\u30c7\u30eb\u304c\u516c\u958b\u3055\u308c\u308b\u306e\u306f\u826f\u3044\u3053\u3068\u3060\u3068\u601d\u3046\u3002\uff08\u6700\u5f8c\u306e\u4f8b\u306f\u60f3\u5b9a\u6d3b\u7528\u4f8b\u304c\u4ed6\u3068\u306f\u3060\u3044\u3076\u7570\u306a\u308a\u305d\u3046\u3067\u306f\u3042\u308b\u304c\u3002\u3002\uff09<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Kimi Linear: An Expressive, Efficient Attention Architecture&nbsp;<\/strong>[75.9]<br>Kimi Linear\u306f\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u306a\u7dda\u5f62\u30a2\u30c6\u30f3\u30b7\u30e7\u30f3\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u3067\u3001\u521d\u3081\u3066\u3001\u516c\u6b63\u306a\u6bd4\u8f03\u3067\u5b8c\u5168\u306b\u30a2\u30c6\u30f3\u30b7\u30e7\u30f3\u3092\u4e0a\u56de\u308a\u307e\u3059\u3002 \u4e2d\u5fc3\u3068\u306a\u308bKimi Delta Attention (KDA)\u306f\u3001Gated DeltaNet\u3092\u62e1\u5f35\u3057\u305f\u8868\u73fe\u529b\u306e\u3042\u308b\u7dda\u5f62\u30a2\u30c6\u30f3\u30b7\u30e7\u30f3\u30e2\u30b8\u30e5\u30fc\u30eb\u3067\u3042\u308b\u3002 \u6211\u3005\u306f,Kimi Linear\u304c\u3088\u308a\u512a\u308c\u305f\u6027\u80fd\u3068\u52b9\u7387\u3067\u5341\u5206\u306a\u6ce8\u610f\u3092\u6255\u3063\u3066,\u30c9\u30ed\u30c3\u30d7\u30a4\u30f3\u3067\u7f6e\u304d\u63db\u3048\u3089\u308c\u308b\u3053\u3068\u3092\u793a\u3059\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2510.26692v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2510.26692v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Thu, 30 Oct 2025 16:59:43 GMT)<\/li>\n\n\n\n<li>\u300cAt its core lies Kimi Delta Attention (KDA), a hardware-efficient linear attention module that extends Gated DeltaNet [111] with a finer-grained gating mechanism. While GDN, similar to Mamba2 [16], employs a coarse head-wise forget gate, KDA introduces a channel-wise variant in which each feature dimension maintains an independent forgetting rate, akin to Gated Linear Attention (GLA) [114]. This fine-grained design enables more precise regulation of the finite-state RNN memory, unlocking the potential of RNN-style models within hybrid architectures.\u300d\u3092\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u69cb\u6210\u3067\u6d3b\u7528\u3002<\/li>\n\n\n\n<li><a href=\"https:\/\/github.com\/MoonshotAI\/Kimi-Linear\">GitHub &#8211; MoonshotAI\/Kimi-Linear<\/a><\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Every Activation Boosted: Scaling General Reasoner to 1 Trillion Open Language Foundation&nbsp;<\/strong>[149.0]<br>Ling 2.0\u306f\u3001\u3059\u3079\u3066\u306e\u30a2\u30af\u30c6\u30a3\u30d9\u30fc\u30b7\u30e7\u30f3\u304c\u63a8\u8ad6\u80fd\u529b\u3092\u4fc3\u9032\u3059\u308b\u3068\u3044\u3046\u539f\u5247\u306b\u57fa\u3065\u3044\u3066\u69cb\u7bc9\u3055\u308c\u305f\u4e00\u9023\u306e\u63a8\u8ad6\u6307\u5411\u306e\u8a00\u8a9e\u57fa\u76e4\u3067\u3042\u308b\u3002 Ling 2.0\u306f\u3001\u7d4c\u9a13\u7684\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u6cd5\u5247\u306b\u3088\u3063\u3066\u5c0e\u304b\u308c\u308b\u3001\u9ad8\u3044\u5206\u6563\u6027\u3001\u30af\u30ed\u30b9\u30b9\u30b1\u30fc\u30eb\u4e00\u8cab\u6027\u3001\u52b9\u7387\u6027\u3092\u5f37\u8abf\u3057\u3066\u3044\u308b\u3002 \u30b7\u30ea\u30fc\u30ba\u306b\u306f\u3001Ling-mini-2.0\u3001Ling-flash-2.0\u3001Ling-1T\u306e3\u3064\u306e\u975e\u601d\u8003\u30e2\u30c7\u30eb\u304c\u542b\u307e\u308c\u3066\u3044\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2510.22115v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2510.22115v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Sat, 25 Oct 2025 01:51:37 GMT)<\/li>\n\n\n\n<li>\u9577\u3044Reasoning\u306b\u30d5\u30a9\u30fc\u30ab\u30b9\u3057\u305f<a href=\"https:\/\/devneko.jp\/wordpress\/?p=7653\">Ring-1T<\/a>\u3068\u306f\u3053\u3068\u306a\u308a\u3001\u4e00\u822c\u7684\u306a\u63a8\u8ad6\u3084\u6307\u793a\u306b\u5f93\u3046\u80fd\u529b\u306b\u30d5\u30a9\u30fc\u30ab\u30b9<\/li>\n\n\n\n<li><a href=\"https:\/\/github.com\/inclusionAI\/Ling-V2\">GitHub &#8211; inclusionAI\/Ling-V2: Ling-V2 is a MoE LLM provided and open-sourced by InclusionAI.<\/a><\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Scaling Latent Reasoning via Looped Language Models&nbsp;<\/strong>[109.6]<br>\u4e8b\u524d\u5b66\u7fd2\u3055\u308c\u305f\u30eb\u30fc\u30d7\u8a00\u8a9e\u30e2\u30c7\u30eb(LoopLM)\u306e\u30d5\u30a1\u30df\u30ea\u30fc\u3067\u3042\u308bOuro\u3092\u63d0\u793a\u3057\u3001\u30aa\u30fc\u30d7\u30f3\u30bd\u30fc\u30b9\u5316\u3059\u308b\u3002 Ouro \u306f (i) \u6f5c\u6642\u7a7a\u9593\u306b\u304a\u3051\u308b\u53cd\u5fa9\u8a08\u7b97, (ii) \u5b66\u7fd2\u6df1\u5ea6\u5272\u308a\u5f53\u3066\u306e\u305f\u3081\u306e\u30a8\u30f3\u30c8\u30ed\u30d4\u30fc\u898f\u5247\u5316\u3055\u308c\u305f\u76ee\u7684, (iii) 7.7T \u30c8\u30fc\u30af\u30f3\u3078\u306e\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u306b\u3088\u308b\u4e8b\u524d\u5b66\u7fd2\u6bb5\u968e\u3078\u306e\u63a8\u8ad6\u3092\u69cb\u7bc9\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2510.25741v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2510.25741v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Wed, 29 Oct 2025 17:45:42 GMT)<\/li>\n\n\n\n<li>Looped Language Model (LoopLM) architecture\u306b\u3088\u308b\u30e2\u30c7\u30eb\u69cb\u7bc9\u306e\u5831\u544a\u3002\u300cwe introduced Ouro, a family of Looped Language Models that demonstrate exceptional parameter efficiency by integrating iterative computation and adaptive depth directly into pre-training on 7.7T tokens. Our 1.4B and 2.6B models consistently match or exceed the performance of 4B and 8B standard transformers, showcasing a 2-3\u00d7 efficiency gain.\u300d\u3068\u975e\u5e38\u306b\u52b9\u7387\u304c\u9ad8\u3044\u3002<\/li>\n\n\n\n<li><a href=\"https:\/\/ouro-llm.github.io\/\">Ouro: Looped Language Models<\/a><\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Parallel Loop Transformer for Efficient Test-Time Computation Scaling\u00a0<\/strong>[34.8]<br>\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb(LLM)\u306f\u5f37\u529b\u3060\u304c\u3001\u63a8\u8ad6\u4e2d\u306b\u73fe\u5b9f\u4e16\u754c\u3067\u4f7f\u3046\u306b\u306f\u9045\u3059\u304e\u308b\u3057\u30b3\u30b9\u30c8\u3082\u304b\u304b\u308b\u3002 \u30eb\u30fc\u30d7\u5909\u63db\u5668\u306f\u3001\u8907\u6570\u306e\u8a08\u7b97\u30b9\u30c6\u30c3\u30d7\u3067\u540c\u3058\u91cd\u307f\u3092\u518d\u5229\u7528\u3059\u308b\u3053\u3068\u3067\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u7bc0\u7d04\u3059\u308b\u3002 \u30eb\u30fc\u30d7\u304c\u6b21\u3005\u3068\u5b9f\u884c\u3055\u308c\u3001\u5404\u8ffd\u52a0\u30eb\u30fc\u30d7\u3067\u63a8\u8ad6\u9045\u5ef6\u3068\u30e1\u30e2\u30ea\u8981\u6c42\u304c\u5897\u5927\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2510.24824v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2510.24824v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Tue, 28 Oct 2025 15:35:50 GMT)<\/li>\n\n\n\n<li>\u3053\u3061\u3089\u306f\u4e26\u5217\u306eParallel Loop Transformer (PLT)<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Emu3.5: Native Multimodal Models are World Learners&nbsp;<\/strong>[65.9]<br>Emu3.5\u306f\u5927\u898f\u6a21\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30eb\u4e16\u754c\u30e2\u30c7\u30eb\u3067\u3001\u8996\u899a\u3068\u8a00\u8a9e\u3092\u307e\u305f\u3044\u3060\u6b21\u306e\u72b6\u614b\u3092\u30cd\u30a4\u30c6\u30a3\u30d6\u306b\u4e88\u6e2c\u3059\u308b\u3002 Emu3.5\u306f\u3001\u8996\u899a\u8a00\u8a9e\u9593\u306e\u30a4\u30f3\u30bf\u30fc\u30ea\u30fc\u30d6\u30c7\u30fc\u30bf\u306e\u30b3\u30fc\u30d1\u30b9\u306b\u57fa\u3065\u3044\u3066\u3001\u4e00\u8cab\u3057\u305f\u6b21\u30c8\u30fc\u30b1\u30f3\u4e88\u6e2c\u76ee\u6a19\u3092\u6301\u3064\u3001\u30a8\u30f3\u30c9\u30c4\u30fc\u30a8\u30f3\u30c9\u3067\u4e8b\u524d\u8a13\u7df4\u3055\u308c\u305f\u3002 \u305d\u308c\u306f\u3001\u4e00\u8cab\u3057\u305f\u4e16\u754c\u63a2\u7d22\u3068\u30aa\u30fc\u30d7\u30f3\u30ef\u30fc\u30eb\u30c9\u306e\u5177\u4f53\u7684\u64cd\u4f5c\u3092\u53ef\u80fd\u306b\u3059\u308b\u3001\u4e00\u822c\u5316\u53ef\u80fd\u306a\u4e16\u754c\u30e2\u30c7\u30ea\u30f3\u30b0\u80fd\u529b\u3092\u793a\u3059\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2510.26583v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2510.26583v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Thu, 30 Oct 2025 15:11:16 GMT)<\/li>\n\n\n\n<li>Emu\u30b7\u30ea\u30fc\u30ba\uff08<a href=\"https:\/\/devneko.jp\/wordpress\/?p=5578\">Emu3: Next-Token Prediction is All You Need \u2013 arXiv\u6700\u65b0\u8ad6\u6587\u306e\u7d39\u4ecb<\/a>\uff09\u306e\u6700\u65b0\u7248\u3002\u300cEmu3.5 further exhibits generalizable worldmodeling abilities encompassing world exploration and embodied manipulation, enabling controllable interaction, free-form navigation, and dynamic scene simulation across both real and imagined environments.  We carefully evaluate these new capabilities and demonstrate clear superiority of Emu3.5, a single 32B unified model, over the closed-source Gemini 2.5 Flash Image [91].\u300d\u3068\u306e\u3053\u3068\u3002<\/li>\n\n\n\n<li><a href=\"https:\/\/emu.world\/pages\/web\/landingPage\">emu.world\/pages\/web\/landingPage<\/a>\u3001<a href=\"https:\/\/github.com\/baaivision\/Emu3.5\">GitHub &#8211; baaivision\/Emu3.5: Native Multimodal Models are World Learners<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u5148\u9031\u306f\u516c\u958b\u30e2\u30c7\u30eb\u306e\u8a71\u984c\u304c\u591a\u304f\u3001\u305d\u306e\u4e2d\u3067\u3082MiniMax-M2 \u3068Kimi-Linear\u306f\u8981\u6ce8\u76ee\u3002\u7279\u306b\u5f8c\u8005\u306f\u52b9\u7387\u6027\u3082\u9ad8\u3044\u3002\u5148\u9031\u306eRing\u3068\u3084\u3084\u3053\u3057\u3044\u304c\u3001Ling-V2\u3082\u5f37\u529b\u306a\u30e2\u30c7\u30eb\u3067\u3042\u308b\uff08This report focus &hellip; <a href=\"https:\/\/devneko.jp\/wordpress\/?p=7682\" class=\"more-link\"><span class=\"screen-reader-text\">&#8220;MiniMax M2, Kimi-Linear,  Ling-V2, Ouro, Emu3.5, gpt-oss-safeguard&#8221; \u306e<\/span>\u7d9a\u304d\u3092\u8aad\u3080<\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[250,293,449,551],"class_list":["post-7682","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-mixture-of-experts","tag-oss","tag-world-model","tag-551"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7682","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7682"}],"version-history":[{"count":2,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7682\/revisions"}],"predecessor-version":[{"id":7714,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7682\/revisions\/7714"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7682"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7682"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7682"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}