Thinking Machines Lab released Inkling on July 15, 2026 — a 975-billion-parameter open-weight AI model that any business can download, modify, and fine-tune on its own data for free. The lab, founded by former OpenAI Chief Technology Officer Mira Murati, states in its own briefing materials that Inkling is “not the strongest model available today, closed or open.”
Inkling is a mixture-of-experts (MoE) model: it holds 975 billion total parameters but activates approximately 41 billion for any given task. Thinking Machines trained it on 45 trillion tokens of text, image, audio, and video, and the model reasons natively across all four modalities. Training ran entirely on Nvidia GB300 NVL72 systems, under a strategic partnership that gives Thinking Machines a gigawatt of Vera Rubin compute.
Why an Open-Weight AI Model Matters for Business Buyers
An open-weight AI model publishes its trained parameters, so a company downloads the model, runs it on its own infrastructure, and fine-tunes it on proprietary data without sending that data to a vendor. Proprietary models such as ChatGPT, Claude, and Gemini expose only an API — the buyer rents capability and keeps none of it.
Bridgewater Associates supplies the clearest proof point. The hedge fund fine-tuned an open-source model on its own financial expertise in partnership with Thinking Machines; the result scored 84.7% on financial reasoning tests at approximately 1/14th the operating cost of top proprietary models, beating them on the benchmark. Thinking Machines and Bridgewater produced that evaluation jointly — no independent party has verified it.
Satya Nadella, Microsoft CEO, warned the same week that enterprises using proprietary AI models pay twice: once in subscription costs, and again by handing over business knowledge embedded in their prompts and corrections, which can be absorbed into future model versions. Microsoft has invested billions in both OpenAI and Anthropic. Clem Delangue, Hugging Face CEO, framed the shift as a division of labour: “Frontier models will increasingly be reserved for experimentation and high-value tasks, while most production AI work shifts to private or open-source alternatives.”
How Thinking Machines Makes Money Without Charging for Inkling
Thinking Machines sells Tinker, its model-customization platform, and gives Inkling away as open weights. Connie Loizos of TechCrunch summarized the strategy: the paid product is the machinery for adapting models, and Inkling is the base model that makes that machinery more valuable. Once weights are public, nothing compels a user to pay Thinking Machines to run the model — revenue depends entirely on the fine-tuning tooling around it.
On a coding benchmark, Inkling uses one-third as many tokens as Nvidia’s Nemotron 3 Ultra to reach the same coding performance. Inkling also lets users dial “thinking effort” up or down to trade reasoning depth for speed, and returns calibrated answers that flag uncertainty rather than guessing.
Where Inkling Sits in the Open-Weight Field
Open-weight frontier models now ship from 5 significant sources, such as Meta (Llama), Tencent (HY3), Moonshot AI (Kimi K2.5), Nvidia (Nemotron), and Thinking Machines (Inkling). Inkling’s differentiator is the explicit enterprise-customization pitch plus the Tinker tooling, not raw benchmark supremacy — a deliberately different position from launches such as GPT-5.6 Sol, where the headline claim is frontier capability. Buyers weighing a hosted proprietary assistant against a self-hosted fine-tune can compare leading AI chatbots on the closed side of that decision before committing infrastructure to the open side.
Thinking Machines reached market in approximately 9 months from founding, against roughly 5 years for OpenAI and 3 years for Anthropic. The company employs approximately 200 people and raised a $2 billion seed round at a $12 billion valuation in 2025. Two co-founders left for OpenAI in January 2026.
- GLM-5.2: Zhipu AI’s 1M-Context Open-Source Model — the open-source release that set the current context-length bar
- Gemma 4 12B: Google’s Multimodal Model That Runs on a Laptop — open weights aimed at local hardware rather than data-centre fine-tuning
- Chinese AI Models Are Running US Enterprise Workloads — how open-weight models entered production stacks
- Anthropic’s $35B Compute Deal With Apollo and Blackstone — the capital scale behind the proprietary alternative
Thinking Machines is the first well-funded lab to say out loud that it is not trying to win on raw benchmarks. That is a $2 billion bet that businesses care more about customization and cost control than about topping a leaderboard. The Bridgewater number — 84.7% at 1/14th the cost — is the whole argument compressed into one line, and it is also the weakest link: it is a joint self-evaluation by the two companies that benefit from it. If an independent lab reproduces something close to it, this becomes the most credible challenge yet to the assumption that the biggest lab wins. Until then, treat Inkling as a serious option to pilot, not a proven replacement.
Frequently Asked Questions
What is Inkling?
Inkling is a 975-billion-parameter open-weight mixture-of-experts AI model released by Thinking Machines Lab on July 15, 2026. It activates approximately 41 billion parameters per task, was trained on 45 trillion tokens of text, image, audio, and video, and is free to download, modify, and fine-tune.
Is Inkling free for business use?
Yes. Thinking Machines publishes Inkling’s weights openly, so companies download and run the model themselves at no licence cost. Thinking Machines earns revenue from Tinker, its paid model-customization platform, rather than from the model itself.
Is Inkling better than ChatGPT, Claude, or Gemini?
No. Thinking Machines states in its own briefing materials that Inkling is “not the strongest model available today, closed or open.” The company’s argument is that a model fine-tuned on a company’s own domain data outperforms a general proprietary model on that company’s tasks, not that Inkling wins general benchmarks.
What is an open-weight AI model?
An open-weight AI model is a model whose trained parameters are published for download, letting any organisation run it on its own infrastructure and fine-tune it on private data. Proprietary models such as ChatGPT and Gemini expose only an API, so the buyer cannot inspect, host, or retrain the underlying model.
