--- title: The New American AI Model Designed to be Customized description: Thinking Machines, founded by ex-OpenAI CTO Mira Murati, released Inkling on July 15, 2026, an open-weight (Apache 2.0) model trained from scratch and designed... canonical: https://daily.dev/posts/the-new-american-ai-model-designed-to-be-customized-qymgfzx81 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: The New American AI Model Designed to be Customized | daily.dev og:description: Thinking Machines, founded by ex-OpenAI CTO Mira Murati, released Inkling on July 15, 2026, an open-weight (Apache 2.0) model trained from scratch and designed... og:url: https://daily.dev/posts/the-new-american-ai-model-designed-to-be-customized-qymgfzx81 og:image: https://api.daily.dev/og/posts/qYmgfzx81.png og:image:alt: The New American AI Model Designed to be Customized og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # The New American AI Model Designed to be Customized **[ByteByteGo](https://daily.dev/sources/bytebytego)** · 21 min read · 1 upvotes · 0 comments ## Summary Thinking Machines, founded by ex-OpenAI CTO Mira Murati, released Inkling on July 15, 2026, an open-weight (Apache 2.0) model trained from scratch and designed for customization. A deep architectural breakdown covers its 975B total / 41B active parameter Mixture-of-Experts design with 256 experts (8 running per token including 2 shared), a bias-based routing balance mechanism, a 5:1 ratio of sliding-window to full-attention layers enabling a 1M-token context, relative position encoding (Shaw-style) chosen over RoPE for better extrapolation, convolution layers for local mixing, an encoder-free multimodal approach using dMel for audio and hMLP patches for images, and a trainable 'effort' setting controlling reasoning length. Trade-offs include high memory floor despite sparse compute, narrow per-layer context, new tooling requirements for the position encoding, and closed training data despite open weights. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How many parameters does Thinking Machines' Inkling model activate per token versus its total size? Inkling stores 975 billion total parameters but activates only about 41 billion per token, roughly 4 percent of the model. This is achieved through a mixture-of-experts design with 256 experts per layer across 66 layers, where a router selects 6 experts per token plus 2 shared experts that run on every token, for 8 experts active in total. _Engineers weighing MoE model hardware costs can track architecture breakdowns like this on daily.dev._ ### Why did Thinking Machines use relative position encoding instead of RoPE for Inkling's one-million-token context window? Relative position encoding, in the style of Shaw and colleagues, was chosen because RoPE's rotation angles only work reliably at positions seen during training, causing problems when extrapolating to unseen lengths like 900,000 tokens. Relative encoding instead learns a value per distance between tokens, so a pair 900,000 tokens apart reuses a distance value seen countless times, avoiding extrapolation entirely. _Developers evaluating long-context model tradeoffs can follow position-encoding design choices via daily.dev._ ### How does Inkling handle images and audio without a separately pretrained encoder? Audio is converted to a mel spectrogram and then quantized into discrete levels using the dMel method, requiring no pretrained audio model. Images are cut into 40x40 pixel patches processed by a small four-stage hMLP stem network, adding under 1% extra compute versus the simplest alternative. Both then pass through a lightweight conversion layer and join text tokens in the same 66 transformer layers, trained from scratch alongside the rest of the model. _Teams building multimodal pipelines can track encoder-free architecture approaches on daily.dev._ ## Similar posts on daily.dev - [ThursdAI - Jul 16 - Inkling 975B open weights, Kimi K3 at 2.8T, a 27B model on a phone & Codex hits 9M](https://daily.dev/posts/thursdai---jul-16---inkling-975b-open-weights-kimi-k3-at-2-8t-a-27b-model-on-a-phone-codex-hits--hledq9pdh) · ThursdAI · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#deep-learning](https://daily.dev/tags/deep-learning), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/the-new-american-ai-model-designed-to-be-customized-qymgfzx81) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"The New American AI Model Designed to be Customized","url":"https://daily.dev/posts/the-new-american-ai-model-designed-to-be-customized-qymgfzx81","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/the-new-american-ai-model-designed-to-be-customized-qymgfzx81"},"datePublished":"2026-08-18T15:34:09.592Z","dateModified":"2026-09-13T19:32:27.371Z","description":"Thinking Machines, founded by ex-OpenAI CTO Mira Murati, released Inkling on July 15, 2026, an open-weight (Apache 2.0) model trained from scratch and designed...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/f14c1f3c6268a1bf3ef7f00c0588134f?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/f14c1f3c6268a1bf3ef7f00c0588134f?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"ByteByteGo","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"ByteByteGo","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/35be29234ee14d01a9cd049c52e12753","url":"https://daily.dev/sources/bytebytego"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/the-new-american-ai-model-designed-to-be-customized-qymgfzx81","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,deep-learning,mixture-of-experts","timeRequired":"PT21M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"ByteByteGo","item":"https://daily.dev/sources/bytebytego"},{"@type":"ListItem","position":3,"name":"The New American AI Model Designed to be Customized"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/the-new-american-ai-model-designed-to-be-customized-qymgfzx81#faq","mainEntity":[{"@type":"Question","name":"How many parameters does Thinking Machines' Inkling model activate per token versus its total size?","acceptedAnswer":{"@type":"Answer","text":"Inkling stores 975 billion total parameters but activates only about 41 billion per token, roughly 4 percent of the model. This is achieved through a mixture-of-experts design with 256 experts per layer across 66 layers, where a router selects 6 experts per token plus 2 shared experts that run on every token, for 8 experts active in total. Engineers weighing MoE model hardware costs can track architecture breakdowns like this on daily.dev."}},{"@type":"Question","name":"Why did Thinking Machines use relative position encoding instead of RoPE for Inkling's one-million-token context window?","acceptedAnswer":{"@type":"Answer","text":"Relative position encoding, in the style of Shaw and colleagues, was chosen because RoPE's rotation angles only work reliably at positions seen during training, causing problems when extrapolating to unseen lengths like 900,000 tokens. Relative encoding instead learns a value per distance between tokens, so a pair 900,000 tokens apart reuses a distance value seen countless times, avoiding extrapolation entirely. Developers evaluating long-context model tradeoffs can follow position-encoding design choices via daily.dev."}},{"@type":"Question","name":"How does Inkling handle images and audio without a separately pretrained encoder?","acceptedAnswer":{"@type":"Answer","text":"Audio is converted to a mel spectrogram and then quantized into discrete levels using the dMel method, requiring no pretrained audio model. Images are cut into 40x40 pixel patches processed by a small four-stage hMLP stem network, adding under 1% extra compute versus the simplest alternative. Both then pass through a lightweight conversion layer and join text tokens in the same 66 transformer layers, trained from scratch alongside the rest of the model. Teams building multimodal pipelines can track encoder-free architecture approaches on daily.dev."}}]} ```