{"id":6561,"date":"2025-04-14T06:32:00","date_gmt":"2025-04-13T21:32:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=6561"},"modified":"2025-04-14T06:32:00","modified_gmt":"2025-04-13T21:32:00","slug":"llama-4-nemotron-h-pangu-ultra-kimi-vl-kimi-vl-thinking","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=6561","title":{"rendered":"Llama 4, Nemotron-H, Pangu Ultra, Kimi-VL, Kimi-VL-Thinking, Deep Coder"},"content":{"rendered":"\n<p>\u5148\u9031\u3082LLM\u95a2\u9023\u306e\u8a71\u984c\u306f\u591a\u304b\u3063\u305f\u304c\u3001Llama4\u306e\u767a\u8868\u306f\u305d\u306e\u4e2d\u3067\u3082\u5927\u304d\u306a\u3082\u306e\u3060\u3063\u305f\uff08<a href=\"https:\/\/ai.meta.com\/blog\/llama-4-multimodal-intelligence\/\">The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation<\/a>\uff09\u3002MoE\u69cb\u6210\u3067\u9ad8\u3044\u6027\u80fd\u3092\u4e3b\u5f35\u3001\u7b2c\u4e09\u8005\u306e\u691c\u8a3c\u3067\u306f\u3044\u307e\u3044\u3061\u3068\u3044\u3046\u8a71\u3082\u3001\u91cf\u5b50\u5316\u306e\u5f71\u97ff\uff08\u6027\u80fd\u52a3\u5316\uff09\u304c\u5927\u304d\u3044\u306e\u3067\u306f\u3068\u3044\u3046\u8a71\u3082\u3042\u3063\u3066\u3001\u691c\u8a3c\u7d50\u679c\u304c\u51fa\u305d\u308d\u3046\u306e\u3092\u5f85\u3061\u305f\u3044\u3068\u3053\u308d\u3002<\/p>\n\n\n\n<p>NVIDIA\u304b\u3089\u306f&nbsp;Mamba-Transformer\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u306aNemotron-H\u304c\u767a\u8868\u3055\u308c\u3066\u3044\u308b\uff08<a href=\"https:\/\/research.nvidia.com\/labs\/adlr\/nemotronh\/\">Nemotron-H: A Family of Accurate, Efficient Hybrid Mamba-Transformer Models &#8211; NVIDIA ADLR<\/a>\uff09\u3002&nbsp;\u300cNemotron-H has been used as the backbone for&nbsp;<a href=\"https:\/\/research.nvidia.com\/publication\/2025-03_cosmos-reason-1-physical-ai-common-sense-embodied-decisions\">Cosmos-Reason 1<\/a>, a very strong VLM for physical AI.\u300d\u3068\u3044\u3046\u306e\u306b\u3082\u6ce8\u76ee\u3002<\/p>\n\n\n\n<p>Huawei\u304b\u3089\u306fPangu Ultra\u306e\u8ad6\u6587\u304c\u51fa\u3066\u3044\u308b\u304c\u3001\u8a73\u7d30\u306aPDF\u306f\u516c\u958b\u3055\u308c\u3066\u3044\u306a\u3044\u3088\u3046\u3002\u300cTo perform such large-scale training efficiently, we utilize 8,192 Ascend NPUs with a series of system optimizations. Evaluations on multiple diverse benchmarks indicate that Pangu Ultra significantly advances the state-of-the-art capabilities of dense LLMs such as Llama 405B and Mistral Large 2, and even achieves competitive results with DeepSeek-R1\u300d\u3068\u3044\u3046\u8208\u5473\u6df1\u3044\u8a18\u8f09\u304c\u3042\u308a\u8a73\u7d30\u304c\u6c17\u306b\u306a\u308b\u3068\u3053\u308d\u3002<\/p>\n\n\n\n<p>Kimi-VL \u306f\u5f37\u529b\u306aMLLM\u3067\u3042\u308a\u3001\u307e\u305f\u3001Kimi-VL-Thinking\u3068LRM\u3067\u3082\u3042\u308b\u306e\u304c\u7279\u5fb4\u7684\u306a\u516c\u958b\u30e2\u30c7\u30eb\uff08<a href=\"https:\/\/huggingface.co\/moonshotai\/Kimi-VL-A3B-Instruct\">moonshotai\/Kimi-VL-A3B-Instruct \u00b7 Hugging Face<\/a>\uff09\u3002o3-mini\u30ec\u30d9\u30eb\u306e\u6027\u80fd\u3092\u4e3b\u5f35\u3059\u308b<a href=\"https:\/\/pretty-radio-b75.notion.site\/DeepCoder-A-Fully-Open-Source-14B-Coder-at-O3-mini-Level-1cf81902c14680b3bee5eb349a512a51\">DeepCoder: A Fully Open-Source 14B Coder at O3-mini Level<\/a>\u306a\u3069\u30aa\u30fc\u30d7\u30f3\u306a\u30e2\u30c7\u30eb\u3082\u9032\u5316\u304c\u901f\u3044\u3002\u30aa\u30fc\u30d7\u30f3\u306a\u30e2\u30c7\u30eb\u3092\u5f37\u5316\u3059\u308b\u65b9\u5411\u3082<a href=\"https:\/\/www.deepcogito.com\/research\/cogito-v1-preview\">Introducing Cogito Preview<\/a>\uff08<a href=\"https:\/\/huggingface.co\/collections\/deepcogito\/cogito-v1-preview-67eb105721081abe4ce2ee53\">Cogito v1 Preview &#8211; a deepcogito Collection<\/a>\uff09\u306a\u3069\u69d8\u3005\u306a\u6210\u679c\u304c\u51fa\u3066\u3044\u3066\u3001\u516c\u958b\u30e2\u30c7\u30eb\u306e\u6027\u80fd\u3082\u5411\u4e0a\u304c\u7d9a\u304f\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Nemotron-H: A Family of Accurate and Efficient Hybrid Mamba-Transformer Models&nbsp;<\/strong>[164.5]<br>\u30cd\u30e2\u30c8\u30ed\u30f3-H\u306f8B\u306856B\/47B\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u30fb\u30de\u30f3\u30d0\u30fb\u30c8\u30e9\u30f3\u30b9\u30d5\u30a9\u30fc\u30de\u30fc\u306e\u30d5\u30a1\u30df\u30ea\u30fc\u3067\u3042\u308b\u3002 \u79c1\u305f\u3061\u306f\u5171\u901a\u306eTransformer\u30e2\u30c7\u30eb\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306b\u304a\u3051\u308b\u307b\u3068\u3093\u3069\u306e\u81ea\u5df1\u6ce8\u610f\u30ec\u30a4\u30e4\u3092Mamba\u30ec\u30a4\u30e4\u306b\u7f6e\u304d\u63db\u3048\u307e\u3059\u3002 Nemotron-H\u30e2\u30c7\u30eb\u306f\u3001\u4ed6\u306e\u540c\u69d8\u306e\u30b5\u30a4\u30ba\u306e\u30aa\u30fc\u30d7\u30f3\u30bd\u30fc\u30b9\u30c8\u30e9\u30f3\u30b9\u30d5\u30a9\u30fc\u30de\u30fc\u30e2\u30c7\u30eb\u3068\u6bd4\u8f03\u3057\u3066\u3001\u7cbe\u5ea6\u304c\u826f\u3044\u304b\u4f4e\u3044\u304b\u306e\u3069\u3061\u3089\u304b\u3092\u63d0\u4f9b\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2504.03624v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2504.03624v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Fri, 04 Apr 2025 17:41:58 GMT)<\/li>\n\n\n\n<li>\u9ad8\u901f\u3001\u9ad8\u6027\u80fd\u306aMamba\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u306aLLM<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Pangu Ultra: Pushing the Limits of Dense Large Language Models on Ascend NPUs&nbsp;<\/strong>[123.3]<br>135\u5104\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u3068\u9ad8\u5bc6\u5ea6\u30c8\u30e9\u30f3\u30b9\u30d5\u30a9\u30fc\u30de\u30fc\u30e2\u30b8\u30e5\u30fc\u30eb\u3092\u6301\u3064\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb(LLM)\u3067\u3042\u308bPangu Ultra\u306b\u3064\u3044\u3066\u8ff0\u3079\u308b\u3002 \u3053\u306e\u3088\u3046\u306a\u5927\u898f\u6a21\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u3092\u52b9\u7387\u7684\u306b\u884c\u3046\u305f\u3081\u306b\u306f,8,192\u500b\u306eAscend NPU\u3068\u4e00\u9023\u306e\u30b7\u30b9\u30c6\u30e0\u6700\u9069\u5316\u3092\u7528\u3044\u308b\u3002 \u6211\u3005\u306e\u8abf\u67fb\u3067\u306f\u3001Ascend NPU\u306f1000\u5104\u4ee5\u4e0a\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u6301\u3064\u9ad8\u5bc6\u5ea6\u30e2\u30c7\u30eb\u3092\u52b9\u7387\u7684\u304b\u3064\u52b9\u679c\u7684\u306b\u8a13\u7df4\u3067\u304d\u308b\u3053\u3068\u3092\u793a\u3057\u305f\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2504.07866v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2504.07866v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Thu, 10 Apr 2025 15:41:51 GMT)<\/li>\n\n\n\n<li>\u30d5\u30a1\u30fc\u30a6\u30a7\u30a4\u306eLLM\u3002\u30d5\u30a1\u30fc\u30a6\u30a7\u30a4\u306e\u30a2\u30af\u30bb\u30e9\u30ec\u30fc\u30bf\u3092\u6d3b\u7528\u3057\u3066\u69cb\u7bc9\u3057\u3066\u3044\u308b\u3068\u306e\u3053\u3068\u3060\u304c\u73fe\u72b6\u8ad6\u6587\u304c\u53c2\u7167\u3067\u304d\u306a\u3044\u72b6\u614b\u3002\u8a73\u7d30\u304c\u6c17\u306b\u306a\u308b\u3068\u3053\u308d\u3002<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Kimi-VL Technical Report&nbsp;<\/strong>[88.1]<br>Kimi-VL\u306f\u8996\u899a\u8a00\u8a9e\u30e2\u30c7\u30eb(VLM)\u3067\u3042\u308a\u3001\u9ad8\u5ea6\u306a\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30eb\u63a8\u8ad6\u3001\u9577\u3044\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u7406\u89e3\u3001\u5f37\u529b\u306a\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u80fd\u529b\u3092\u63d0\u4f9b\u3059\u308b\u3002 \u6c4e\u7528 VLM \u3068\u3057\u3066\u3001Kimi-VL \u306f\u30de\u30eb\u30c1\u30bf\u30fc\u30f3\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u30bf\u30b9\u30af(OSWorld \u306a\u3069)\u306b\u512a\u308c\u3001\u65d7\u8266\u30e2\u30c7\u30eb\u3068\u4e00\u81f4\u3059\u308b\u3002 Kimi-VL\u3092\u30d9\u30fc\u30b9\u3068\u3057\u3066\u3001Kim-VL-Thinking\u3068\u3044\u3046\u5148\u9032\u7684\u306a\u30ed\u30f3\u30b0\u30b7\u30f3\u30ad\u30f3\u30b0\u30e2\u30c7\u30eb\u3092\u5c0e\u5165\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2504.07491v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2504.07491v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Thu, 10 Apr 2025 06:48:26 GMT)<\/li>\n\n\n\n<li>\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u30bf\u30b9\u30af\u3067\u3082\u9ad8\u3044\u6027\u80fd\u3092\u6301\u3064\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30ebLLM\u3002Thinking\u30d0\u30fc\u30b8\u30e7\u30f3\u306f\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u3068\u6bd4\u8f03\u3057\u3066\u9ad8\u3044\u6027\u80fd\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/github.com\/MoonshotAI\/Kimi-VL\">GitHub &#8211; MoonshotAI\/Kimi-VL: Kimi-VL: Mixture-of-Experts Vision-Language Model for Multimodal Reasoning, Long-Context Understanding, and Strong Agent Capabilities<\/a>, <a href=\"https:\/\/huggingface.co\/moonshotai\/Kimi-VL-A3B-Instruct\">moonshotai\/Kimi-VL-A3B-Instruct \u00b7 Hugging Face<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u5148\u9031\u3082LLM\u95a2\u9023\u306e\u8a71\u984c\u306f\u591a\u304b\u3063\u305f\u304c\u3001Llama4\u306e\u767a\u8868\u306f\u305d\u306e\u4e2d\u3067\u3082\u5927\u304d\u306a\u3082\u306e\u3060\u3063\u305f\uff08The Llama 4 herd: The beginning of a new era of natively multimodal  &hellip; <a href=\"https:\/\/devneko.jp\/wordpress\/?p=6561\" class=\"more-link\"><span class=\"screen-reader-text\">&#8220;Llama 4, Nemotron-H, Pangu Ultra, Kimi-VL, Kimi-VL-Thinking, Deep Coder&#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":[223,232,235,250],"class_list":["post-6561","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-llm","tag-lrm","tag-mamba","tag-mixture-of-experts"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/6561","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=6561"}],"version-history":[{"count":0,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/6561\/revisions"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6561"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6561"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6561"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}