{"id":11607,"date":"2026-11-21T20:30:24","date_gmt":"2026-11-21T12:30:24","guid":{"rendered":"https:\/\/goutanddesign.com\/wpress\/?p=11607"},"modified":"2026-11-21T20:30:24","modified_gmt":"2026-11-21T12:30:24","slug":"inkling-2-mini-a-looping-open-model-that-rethinks-ai-reasoning","status":"publish","type":"post","link":"https:\/\/goutanddesign.com\/wpress\/2026\/11\/inkling-2-mini-a-looping-open-model-that-rethinks-ai-reasoning\/","title":{"rendered":"Inkling 2 Mini: A Looping Open Model That Rethinks AI Reasoning"},"content":{"rendered":"<span class=\"span-reading-time rt-reading-time\" style=\"display: block;\"><span class=\"rt-label rt-prefix\">Reading Time : <\/span> <span class=\"rt-time\"> 3<\/span> <span class=\"rt-label rt-postfix\">minutes<\/span><\/span><p style=\"text-align: justify;\">Thinking Machines Lab has released Inkling 2 Mini, an open-weight model that raises an uncomfortable question for anyone who assumed AI reasoning had to appear in visible text. The model is small enough to run on a laptop, yet it appears to reason by repeatedly passing information through the same part of its network. Instead of writing every step into a chain of thought, it may do much of that work internally. The release, which follows Inkling 2 Small, invites a closer look at whether efficiency and transparency can coexist in a model designed for autonomous work.<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-11608\" src=\"https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z2-2.webp\" alt=\"\" width=\"1000\" height=\"563\" srcset=\"https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z2-2.webp 1000w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z2-2-300x169.webp 300w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z2-2-768x432.webp 768w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z2-2-600x338.webp 600w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p style=\"text-align: justify;\">The technical trick is deceptively simple. Inkling 2 Mini has 24 layers, but its middle 12 run three times for every token, while the first and last six run once. That produces 48 layer applications from only 24 physical layers, because the repeated middle layers share the same weights. The result is a mixture-of-experts model with 26.4 billion total parameters and 3.4 billion active parameters, or 5.4 billion weight activations per token when the loops are counted. Most weights were trained and deployed natively in NVFP4, with the final four layers in MXFP8, keeping the full footprint near 20GB.<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-11609\" src=\"https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z3-3.png\" alt=\"\" width=\"1000\" height=\"525\" srcset=\"https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z3-3.png 1000w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z3-3-300x158.png 300w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z3-3-768x403.png 768w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z3-3-600x315.png 600w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p style=\"text-align: justify;\">The performance claims make the design more than a curiosity. Thinking Machines says Inkling 2 Mini leads every open-weight model under 100 billion parameters on DeepSWE v1.1, at 48.2%, and on Vals Index v2.2, at 46.5%. Artificial Analysis scored it at 20 on its Intelligence Index, matching Qwen3.8 27B while using roughly one-fifth as many active parameters. The company also claims the model is about six times cheaper and four times faster per task than that comparison model, though its looped architecture uses about 38% more output tokens for the same score while consuming roughly 2.8 times less compute.<\/p>\n<p>&nbsp;<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-11610\" src=\"https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z4-11.jpg\" alt=\"\" width=\"1000\" height=\"525\" srcset=\"https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z4-11.jpg 1000w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z4-11-300x158.jpg 300w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z4-11-768x403.jpg 768w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2026\/11\/z4-11-600x315.jpg 600w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p style=\"text-align: justify;\">The caveats deserve as much attention as the benchmarks. The model scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, trailing Qwen3.8-27B on both multi-turn conversation and instructions containing many constraints. Its own model card warns that it may take destructive actions, such as deleting or overwriting files, without confirmation, and recommends running agentic workloads in a sandbox. More broadly, looping could make a model harder to audit because computation happens without producing readable reasoning. Thinking Machines says its chain of thought becomes shorter but remains legible at this scale, and it has released a variable-loop research model so independent researchers can investigate the claim. The weights are available under Apache 2.0, with support from SGLang, vLLM, Ollama, NVIDIA Dynamo and other inference tools.<\/p>\n<p>&nbsp;<\/p>\n<figure id=\"attachment_8029\" aria-describedby=\"caption-attachment-8029\" style=\"width: 1000px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" class=\"size-full wp-image-8029\" src=\"https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2025\/06\/Ben-footer-jul2024.jpg\" alt=\"B\u00e9n\u00e9dicte Lin \u2013 Brussels, Paris, London, Beijing, Seoul, Bangkok, Tokyo, New York, Taipei, Hong Kong\" width=\"1000\" height=\"379\" srcset=\"https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2025\/06\/Ben-footer-jul2024.jpg 1000w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2025\/06\/Ben-footer-jul2024-600x227.jpg 600w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2025\/06\/Ben-footer-jul2024-300x114.jpg 300w, https:\/\/goutanddesign.com\/wpress\/wp-content\/uploads\/2025\/06\/Ben-footer-jul2024-768x291.jpg 768w\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" \/><figcaption id=\"caption-attachment-8029\" class=\"wp-caption-text\">B\u00e9n\u00e9dicte Lin \u2013 Brussels, Paris, London, Beijing, Seoul, Bangkok, Tokyo, New York, Taipei, Hong Kong<\/figcaption><\/figure>\n<p>&nbsp;<\/p>\n<p>#AI #Inkling2Mini #ThinkingMachines #OpenWeights #LoopedTransformer #MachineLearning #AgenticAI #LocalAI #AISafety #AIResearch<\/p>\n","protected":false},"excerpt":{"rendered":"<p><span class=\"span-reading-time rt-reading-time\" style=\"display: block;\"><span class=\"rt-label rt-prefix\">Reading Time : <\/span> <span class=\"rt-time\"> 3<\/span> <span class=\"rt-label rt-postfix\">minutes<\/span><\/span>Inkling 2 Mini is an open-weight model that runs its middle layers three times for every token, gaining internal computation without expanding its visible reasoning. It is small enough for a laptop, strong at agentic coding, and released under Apache 2.0. But its looping design also raises a difficult question: if a model reasons inside its network, how much of that process can outsiders actually inspect?<\/p>\n","protected":false},"author":2,"featured_media":11611,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[26,14,374,752,483],"tags":[589,207,206,19,27,217,15,32,4,838,93],"class_list":["post-11607","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-branding","category-e-commerce","category-economy","category-fintech","category-technology","tag-589","tag-ai","tag-artificial-intelligence","tag-benedicte","tag-branding","tag-china","tag-e-commerce","tag-entreprise-strategy","tag-gout-design","tag-inkling","tag-investments"],"_links":{"self":[{"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/posts\/11607","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/comments?post=11607"}],"version-history":[{"count":1,"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/posts\/11607\/revisions"}],"predecessor-version":[{"id":11612,"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/posts\/11607\/revisions\/11612"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/media\/11611"}],"wp:attachment":[{"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/media?parent=11607"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/categories?post=11607"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/goutanddesign.com\/wpress\/wp-json\/wp\/v2\/tags?post=11607"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}