--- title: Mira Murati's first model is a bet that customizable AI... description: Mira Murati's new AI startup Thinking Machines Lab has released Inkling, a 975B parameter mixture-of-experts open-weight model trained on 45 trillion... canonical: https://daily.dev/posts/mira-murati-s-first-model-is-a-bet-that-customizable-ai-beats-centralized-ai-hbolh6dmv twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Mira Murati's first model is a bet that customizable AI beats centralized AI | daily.dev og:description: Mira Murati's new AI startup Thinking Machines Lab has released Inkling, a 975B parameter mixture-of-experts open-weight model trained on 45 trillion... og:url: https://daily.dev/posts/mira-murati-s-first-model-is-a-bet-that-customizable-ai-beats-centralized-ai-hbolh6dmv og:image: https://api.daily.dev/og/posts/hBOLH6dMv.png og:image:alt: Mira Murati's first model is a bet that customizable AI beats centralized AI 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. # Mira Murati's first model is a bet that customizable AI beats centralized AI **[Trends](https://daily.dev/sources/trends)** · 5 min read · 3 upvotes · 0 comments ## Summary Mira Murati's new AI startup Thinking Machines Lab has released Inkling, a 975B parameter mixture-of-experts open-weight model trained on 45 trillion multimodal tokens with a 1M token context window. The model supports text, image, and audio inputs and is hosted on Together AI. The real thesis isn't the model itself but the company's philosophical bet that customizable, distributed AI beats centralized models controlled by a single lab. Paired with a fine-tuning platform called Tinker, the business model centers on enterprise customization. A Bridgewater Associates case study shows a fine-tuned open model outperformed top proprietary models on financial reasoning at 1/14th the cost. The argument draws on Hayek's dispersed knowledge theory and critiques current AI benchmarks for measuring autonomous task completion rather than human-AI collaboration. One caveat: Inkling used partial distillation from other open-weight models for early training, and the 'customizable AI wins' thesis is also conveniently aligned with not yet being able to compete on raw capability with OpenAI or Anthropic. ## Content Thinking Machines Lab just shipped Inkling, its first open-weights model, and the reaction is roughly what you'd expect when a $12B company releases something that's good but not obviously the best: a mix of genuine enthusiasm, measured skepticism, and a few people pointing out the clock is already ticking. The basics: Inkling is a 975B-parameter mixture-of-experts model (41B active per token), trained on 45 trillion multimodal tokens, with a 1M token context window and native support for text, image, and audio inputs. Weights are on Hugging Face under Apache license. Day-zero support landed across vLLM, Together AI, Modal, Databricks, and Baseten. Unsloth already has a 1-bit GGUF quant running at 30-40 tokens/second that shrinks the 1.9TB model down to around 270GB. Soumith Chintala called it "yours to personalize and use openly." Lilian Weng framed it as "a foundation with solid performance across a broad range of capabilities." The team is clearly proud, and they should be — building a near-trillion-parameter multimodal model from scratch in roughly nine months is not nothing. ## The benchmark picture is complicated Here's where it gets interesting. The company is upfront that Inkling doesn't top every leaderboard. On the Artificial Analysis Intelligence Index it scores 41, behind DeepSeek V4 Pro and MiniMax M3 at 44, Kimi K2.6 at 43, and MiMo V2.5 Pro at 42. One account called it "basically another Kimi-K2.6" and suggested the release was timed to get ahead of Kimi-K3 and DeepSeek-V4-GA dropping. That same account later walked it back slightly, saying Inkling is "the smartest open-weight model right now" — but added it probably has 1-2 weeks before Kimi-K3 weights arrive. Ethan Mollick was more direct: "Happy to see a new open weights model, but, so far, Inkling is pretty rough in my tests, not close to frontier Chinese open weights models." The numbers back up some of that skepticism. SWE-bench Verified is 77.6% (the strongest open comparator hits ~80.6%, best closed is 95%). Tau 3 Banking — which tests multi-step transaction logic and policy adherence — is 23.7%, which is genuinely weak for anything you'd want running unsupervised in a production workflow. SimpleQA is 43.9%, suggesting you shouldn't treat this as a reliable closed-book factual engine. On the other hand: 91.4% on VoiceBench is legitimately strong for native audio understanding. Design Arena Elo of 1,257 puts it right next to GPT-5.6 Sol (1,260) for web dev. ARC-AGI-2 at 36.5% is respectable. And the architecture has some genuinely interesting choices — relative position bias instead of RoPE, short convolutions in several places, a 5:1 local-to-global attention ratio — that caught the attention of people like Sebastian Raschka. ## The real product might be Inkling-Small Here's the thing that's easy to miss in the headline numbers: Inkling-Small, a preview model with 276B total parameters and 12B active, nearly matches the flagship on several benchmarks. Humanity's Last Exam: 29.6% vs 29.7%. GPQA: 88.3% vs 87.2%. SWE-bench Verified: 77.4% vs 77.6%. The gaps open up on harder tasks — Terminal Bench drops from 63.8% to 52.7%, Tau 3 Banking falls from 23.7% to 13.6%, SimpleQA collapses from 43.9% to 20.9%. So the flagship's extra capacity is mostly showing up in factual coverage and difficult long-horizon agent tasks, not in the reasoning and coding benchmarks that get the most attention. If those Small numbers hold up under independent testing, a lot of workloads will rationally skip the flagship entirely. The smaller model has a much more manageable deployment footprint and could unlock self-hosted adoption for organizations that can't spin up 16 H200s. ## The actual business model Inkling is free. The money is in Tinker, the company's managed fine-tuning platform. The pitch: take Inkling (or honestly, any open model — Tinker already ran a Bridgewater case study on Qwen3-235B), add your proprietary data and expert labels, and get a specialized model without managing distributed training infrastructure yourself. The Bridgewater example is the most concrete evidence for this thesis. A fine-tuned model on financial article classification reportedly hit 84.7% accuracy versus 78.2% for the best frontier model with prompting, at 13.8x lower inference cost. That's a real result, though it comes with the usual caveats: vendor-reported, narrow task, not independently replicated, and it used a third-party base model. The model-agnosticism of Tinker is both a strength and a tell. It means the platform can win regardless of who has the best base model this month. It also means Thinking Machines isn't betting the company on Inkling staying competitive — which is probably wise given the pace of open-weight releases right now. One genuinely interesting demo: Inkling fine-tuning itself. The model generated a fine-tuning job, produced synthetic training examples, kicked off training on Tinker, loaded the resulting weights, and validated the behavior change — in about 27 minutes. The task was trivial (a lipogram constraint), but the loop it demonstrates is the actual product vision. ## What to make of it Inkling is a credible first release from a team that clearly knows what it's doing. The architecture is thoughtful, the multimodal coverage is real, and the open-weights commitment with Apache licensing is meaningful. The ecosystem picked it up fast — vLLM, Unsloth, Modal, Databricks all had day-zero support. But "credible" and "best" aren't the same thing, and the company knows it. The explicit positioning as a customizable foundation rather than a benchmark leader is either honest self-assessment or a smart way to frame a mid-table result, depending on how charitable you're feeling. Probably both. The harder question is whether Tinker can build durable value before the base-model market commoditizes further. The Bridgewater case is promising. The Tau 3 Banking score is a reminder that "agentic" on a benchmark and "reliable in production" are still pretty different things. --- 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/mira-murati-s-first-model-is-a-bet-that-customizable-ai-beats-centralized-ai-hbolh6dmv) ```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":"Mira Murati's first model is a bet that customizable AI beats centralized AI","url":"https://daily.dev/posts/mira-murati-s-first-model-is-a-bet-that-customizable-ai-beats-centralized-ai-hbolh6dmv","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/mira-murati-s-first-model-is-a-bet-that-customizable-ai-beats-centralized-ai-hbolh6dmv"},"datePublished":"2026-07-15T18:07:30.298Z","dateModified":"2026-07-20T17:08:01.020Z","description":"Mira Murati's new AI startup Thinking Machines Lab has released Inkling, a 975B parameter mixture-of-experts open-weight model trained on 45 trillion...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/3da6ac5de49b243fe13991f6652db4e4?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/3da6ac5de49b243fe13991f6652db4e4?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Trends","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":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/mira-murati-s-first-model-is-a-bet-that-customizable-ai-beats-centralized-ai-hbolh6dmv","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,deep-learning,mixture-of-experts","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Mira Murati's first model is a bet that customizable AI beats centralized AI"}]} ```