{"id":10173,"date":"2026-07-13T18:02:20","date_gmt":"2026-07-13T10:02:20","guid":{"rendered":"https:\/\/infernews.com\/blog\/26m-function-call-model-that-runs-on-incredibly-small-devices\/"},"modified":"2026-07-13T18:03:35","modified_gmt":"2026-07-13T10:03:35","slug":"26m-function-call-model-that-runs-on-incredibly-small-devices","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/26m-function-call-model-that-runs-on-incredibly-small-devices\/","title":{"rendered":"Needle \u60f3\u628a\u5fae\u578b AI \u5e36\u843d\u624b\u6a5f\u540c\u624b\u9336"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/banner.jpg\" alt=\"Logo\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u60f3\u55ba\u624b\u6a5f\u3001\u624b\u9336\u6216\u8005\u773c\u93e1\u4e00\u985e\u88dd\u7f6e\u653e\u5165\u53ef\u7528\u5605\u500b\u4eba AI\uff0c\u5361\u4f4d\u5f80\u5f80\u5514\u4fc2\u6a21\u578b\u5920\u5514\u5920\u5927\uff0c\u800c\u4fc2\u5920\u5514\u5920\u7d30\u3001\u5920\u5514\u5920\u5feb\uff0c\u4ef2\u8981\u80af\u505a\u5de5\u5177\u547c\u53eb\u3002Needle \u5c31\u4fc2\u671d\u5462\u500b\u4f4d\u7f6e\u843d\u624b\uff1a\u4e00\u500b\u4ee5 <strong>Simple Attention Network<\/strong> \u70ba\u6838\u5fc3\u5605\u5fae\u578b\u6a21\u578b\u9805\u76ee\uff0c\u91cd\u9ede\u8655\u7406 single-shot function call\uff0c\u76ee\u6a19\u5514\u4fc2\u9577\u7bc7\u5c0d\u8a71\uff0c\u800c\u4fc2\u5e6b\u500b\u4eba AI \u66f4\u7a69\u5b9a\u5481\u53eb\u5de5\u5177\u505a\u4e8b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5462\u500b\u9805\u76ee\u6700\u503c\u5f97\u7559\u610f\u5605\u5730\u65b9\uff0c\u5728\u65bc\u4f62\u5c07 Gemini 3.1 \u84b8\u993e\u5230 26M \u53c3\u6578\uff0c\u4e26\u4e14\u4fdd\u7559\u5230\u53ef\u4ee5\u55ba Mac\/PC \u672c\u5730 finetune \u5605\u8def\u7dda\u3002\u5c0d\u958b\u767c\u8005\u540c\u7522\u54c1\u5718\u968a\u569f\u8b1b\uff0c\u610f\u601d\u597d\u76f4\u63a5\uff1a\u4f60\u672a\u5fc5\u8981\u7d81\u6b7b\u96f2\u7aef\u5927\u6a21\u578b\uff0c\u4ea6\u53ef\u4ee5\u5148\u7528\u958b\u653e\u6b0a\u91cd\u540c\u8cc7\u6599\u751f\u6210\u6d41\u7a0b\uff0c\u8a66\u81ea\u5df1\u5605\u5de5\u5177\u4ecb\u9762\u3001\u6307\u4ee4\u683c\u5f0f\u540c function schema\uff0c\u518d\u6309\u9700\u8981\u5fae\u8abf\u3002<\/p>\n\n\n<figure class=\"wp-block-embed-youtube wp-block-embed is-type-video is-provider-youtube wp-embed-aspect-16-9 wp-has-aspect-ratio\"><div class=\"lyte-wrapper\" title=\"Cactus Needle - The 26M Function Calling Model\" style=\"width:853px;max-width:100%;margin:5px auto;\"><div class=\"lyMe\" id=\"WYL_tt9UJ0NiOzU\" itemprop=\"video\" itemscope itemtype=\"https:\/\/schema.org\/VideoObject\"><div><meta itemprop=\"thumbnailUrl\" content=\"https:\/\/infernews.com\/blog\/wp-content\/plugins\/wp-youtube-lyte\/lyteCache.php?origThumbUrl=https%3A%2F%2Fi.ytimg.com%2Fvi%2Ftt9UJ0NiOzU%2Fhqdefault.jpg\" \/><meta itemprop=\"embedURL\" content=\"https:\/\/www.youtube.com\/embed\/tt9UJ0NiOzU\" \/><meta itemprop=\"duration\" content=\"PT14M30S\" \/><meta itemprop=\"uploadDate\" content=\"2026-07-12T13:00:18Z\" \/><\/div><meta itemprop=\"accessibilityFeature\" content=\"captions\" \/><div id=\"lyte_tt9UJ0NiOzU\" data-src=\"https:\/\/infernews.com\/blog\/wp-content\/plugins\/wp-youtube-lyte\/lyteCache.php?origThumbUrl=https%3A%2F%2Fi.ytimg.com%2Fvi%2Ftt9UJ0NiOzU%2Fhqdefault.jpg\" class=\"pL\"><div class=\"tC\"><div class=\"tT\" itemprop=\"name\">Cactus Needle - The 26M Function Calling Model<\/div><\/div><div class=\"play\"><\/div><div class=\"ctrl\"><div class=\"Lctrl\"><\/div><div class=\"Rctrl\"><\/div><\/div><\/div><noscript><a href=\"https:\/\/youtu.be\/tt9UJ0NiOzU\" rel=\"nofollow noopener\" target=\"_blank\"><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/plugins\/wp-youtube-lyte\/lyteCache.php?origThumbUrl=https%3A%2F%2Fi.ytimg.com%2Fvi%2Ftt9UJ0NiOzU%2F0.jpg\" alt=\"Cactus Needle - The 26M Function Calling Model\" width=\"853\" height=\"460\" \/><br \/>Watch this video on YouTube<\/a><\/noscript><meta itemprop=\"description\" content=\"In this video, I look at Needle by the company Cactus. This is a 26 million parameter function calling model with a very unique architecture that is outperforming a lot of models much bigger than itself. Blog: https:\/\/cactuscompute.com\/blog\/needle Demo: https:\/\/huggingface.co\/spaces\/shreyask\/needle-playground Github: https:\/\/github.com\/cactus-compute\/needle Twitter: https:\/\/x.com\/Sam_Witteveen \ud83d\udd75\ufe0f Interested in building LLM Agents? Fill out the form below Building LLM Agents Form: https:\/\/drp.li\/dIMes \ud83d\udc68\u200d\ud83d\udcbbGithub: https:\/\/github.com\/samwit\/llm-tutorials \u23f1\ufe0fTime Stamps: 00:00 Intro 01:00 Cactus Needle 01:21 Blog 03:39 Baseline Transformer 04:00 Architecture 05:23 GitHub Repo 07:00 Demo\"><\/div><\/div><div class=\"lL\" style=\"max-width:100%;width:853px;margin:5px auto;\"><\/div><figcaption><\/figcaption><\/figure>\n\n\n<p class=\"wp-block-paragraph\">\u540c\u985e\u5c0f\u6a21\u578b\u901a\u5e38\u6703\u55ba\u300c\u5c3a\u5bf8\u3001\u901f\u5ea6\u3001\u6cdb\u5316\u80fd\u529b\u300d\u4e4b\u9593\u62c9\u626f\uff0cNeedle \u660e\u986f\u63c0\u5497\u529f\u80fd\u5c0e\u5411\u5462\u4e00\u908a\u3002README \u5df2\u7d93\u8b1b\u5f97\u5f88\u5766\u767d\uff1a\u4f62\u55ba single-shot function call \u52dd\u904e FunctionGemma-270m\u3001Qwen-0.6B\u3001Graninte-350m\u3001LFM2.5-350m\uff0c\u4f46\u5462\u985e\u8f03\u5927\u6a21\u578b\u55ba\u5c0d\u8a71\u7bc4\u570d\u540c\u5bb9\u91cf\u4e0a\u4ecd\u7136\u66f4\u5f37\uff0c\u6240\u4ee5 Needle \u6bd4\u8f03\u4f3c\u4e00\u628a\u5c08\u7528\u5de5\u5177\uff0c\u800c\u5514\u4fc2\u901a\u624d\u52a9\u624b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u985e\u578b\u4e0a\u5c6c\u65bc\u958b\u6e90\u6a21\u578b\u9805\u76ee\uff0c\u96c6\u4e2d\u89e3\u6c7a\u5c0f\u88dd\u7f6e\u4e0a\u5605 function call \u6548\u7387\u8207\u90e8\u7f72\u6210\u672c\u3002<\/li>\n\n\n\n<li>\u6b0a\u91cd\u540c dataset generation \u90fd\u5df2\u958b\u653e\uff0c\u9069\u5408\u62ff\u4f86\u6e2c\u8a66\u81ea\u5bb6\u5de5\u5177\u93c8\u540c\u5fae\u8abf\u6d41\u7a0b\u3002<\/li>\n\n\n\n<li>\u751f\u7522\u74b0\u5883\u914d\u5408 Cactus\uff0c\u53ef\u9054 6000 toks\/sec prefill \u540c 1200 decode speed\uff0c\u53d6\u5411\u975e\u5e38\u8457\u91cd\u541e\u5410\u3002<\/li>\n\n\n\n<li>\u9810\u8a13\u7df4\u7528 16 TPU v6e \u8dd1 200B tokens\uff0c\u4e4b\u5f8c\u518d\u7528 2B tokens \u5605 single-shot function call dataset \u505a post-training\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6a21\u578b\u7d50\u69cb\u4ea6\u53cd\u6620\u5497\u5462\u7a2e\u53d6\u5411\uff1aSimple Attention Network \u63a1\u7528 encoder-decoder \u4f48\u5c40\uff0c\u914d\u5408 GQA+RoPE\u3001Cross Attn\u3001ZCRMSNorm \u540c shared embedding\uff0c\u76ee\u7684\u4fc2\u7528\u66f4\u7d30\u898f\u6a21\u652f\u6490\u5de5\u5177\u547c\u53eb\u8f38\u51fa\u3002\u8981\u7559\u610f\u5605\u9650\u5236\u540c\u6a23\u6e05\u695a\uff0c\u5c0f\u6a21\u578b\u672c\u8eab\u6bd4\u8f03 finicky\uff0c\u5c0d\u8cc7\u6599\u683c\u5f0f\u3001\u5de5\u5177\u5b9a\u7fa9\u540c\u5fae\u8abf\u8cea\u7d20\u6703\u66f4\u654f\u611f\uff1b\u9700\u8981\u7a69\u5b9a\u591a\u8f2a\u5c0d\u8a71\u6216\u8005\u66f4\u5ee3\u77e5\u8b58\u8986\u84cb\u5605\u5834\u666f\uff0c\u4ecd\u7136\u672a\u5fc5\u4fc2 Needle \u6700\u5408\u9069\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/cactus-compute\/needle\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Needle\u5514\u4fc2\u8ffd\u6c42\u842c\u80fd\u5c0d\u8a71\uff0c\u800c\u4fc2\u96c6\u4e2d\u505a\u597d function call\u300226M \u53c3\u6578\u505a\u5230\u53ef\u672c\u5730\u5fae\u8abf\uff0c\u65b9\u5411\u76f8\u7576\u9bae\u660e\u3002<\/p>\n","protected":false},"author":8,"featured_media":10172,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,199,131,140,166,185,76,127],"tags":[],"class_list":["post-10173","post","type-post","status-publish","format-standard","hentry","category-133","category-dataset-","category-embedding","category-gemini","category-mac","category-qwen","category-76","category-127"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10173","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=10173"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10173\/revisions"}],"predecessor-version":[{"id":10175,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10173\/revisions\/10175"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10172"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10173"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10173"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10173"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}