Mira Murati's Inkling Challenges OpenAI's API Dominance Strategy

Thinking Machines Inkling and the Rise of Massive Open Models Let's talk about Inkling, because the way Thinking Machines launched this model tells you almost as much as the model itself. **Technical Deep Dive** Here's what we're actually looking at: Inkling is a Mixture-of-Experts transformer with 975 billion total parameters, but only 41 billion active at any given time. That distinction matters enormously.

MoE architecture means the model routes each input through a specialized subset of its total capacity — so you get near-frontier scale without the compute cost of running 975 billion parameters on every single token. It was pretrained on 45 trillion tokens spanning text, images, audio, and video, and it supports a one-million-token context window, which is genuinely enormous for an open-weight release. Alongside the full model, Thinking Machines is also previewing Inkling-Small — 12 billion active parameters, same training recipe, optimized for cost and latency.

That two-tier release is smart product strategy: one model for organizations that want to fine-tune a capable foundation, another for teams that need something deployable tomorrow without a GPU cluster. The model is natively multimodal — it reasons across text, images, and audio out of the box — and it ships with controllable thinking effort, meaning you can dial down reasoning depth to cut costs when you don't need heavy computation. On one coding benchmark, Inkling reportedly needed only a third of the tokens Nvidia's Nemotron required to match its score.

That's an efficiency argument, not a capability argument, and Thinking Machines knows the difference. **Financial Analysis** Thinking Machines built all of this in roughly nine months. That timeline is striking, and it frames the financial story here: this is a company betting that the economics of frontier model development are unsustainable for most buyers, and that the real money is in the customization layer on top of a solid open foundation.

The business model is clear. Inkling ships free and open-weight on Hugging Face. Revenue flows through Tinker, Thinking Machines' cloud-based fine-tuning platform, where companies pay to build specialized versions of the model on their own data.

They cite one example: hedge fund Bridgewater reportedly fine-tuned an open model on its proprietary expertise and hit 84.7% on financial reasoning benchmarks at one-fourteenth the cost of equivalent closed models. That's the pitch to enterprise buyers in a single sentence.

If you have proprietary data and an ML team, the total cost of ownership for a customized open model could be dramatically lower than paying per-token API fees to OpenAI or Anthropic indefinitely. The catch, and it's a real one, is that "if you have an ML team" is doing a lot of heavy lifting in that sentence. Most enterprises don't.

**Market Disruption** Here's the competitive dynamic worth understanding. The open-weight model landscape has been dominated by Chinese labs — DeepSeek, GLM from Zhipu, and others — since Meta began pulling back from its fully open-source commitments. The Rundown noted explicitly that Inkling gives the U.

S. a rare homegrown open-source player in a lane Chinese labs have owned. That framing is geopolitically loaded and probably intentional.

But let's be honest about where Inkling actually sits in the market. Ethan Mollick, Wharton professor and one of the more rigorous AI benchmarkers out there, tested the model and said it isn't close to leading open-weight Chinese models. Ben's Bites echoed that assessment.

So this is not a model that threatens GLM-5.2 or the top DeepSeek releases on raw capability today. What it does do is establish Thinking Machines as a credible player in the open-weight space and give TechCrunch and Wired something to write about — which is itself a market signal.

Two months ago, Thinking Machines previewed interactive collaboration models. Now they have a foundation model. The roadmap is becoming visible.

The deeper disruption may be to the closed-API business model itself. Every capable open-weight release puts pressure on OpenAI and Anthropic to justify their per-token pricing. When companies can fine-tune a 41-billion-active-parameter multimodal model on their own infrastructure, the "just call our API" pitch gets harder to make to cost-conscious buyers.

**Cultural & Social Impact** Mira Murati's name carries real weight here, and Thinking Machines is leaning into it. As former CTO of OpenAI, Murati was present for GPT-4, the ChatGPT launch, and the period when AI went from research curiosity to cultural phenomenon. Her departure from OpenAI in late 2024 was widely seen as a signal that something had shifted inside that organization.

Inkling's positioning — explicitly not the most powerful model, explicitly built for customization, explicitly open-weight — reads as a philosophical statement as much as a product launch. It's an argument that the frontier race is the wrong race, that AI development should be inspectable and adaptable rather than locked behind API walls. Whether that resonates culturally depends on who's listening.

Developers and researchers will respond to it. Enterprise buyers who want a vendor relationship and don't want to manage infrastructure probably won't. The model of safety responsibility shifting to the customer when you release open weights is also worth naming clearly: Thinking Machines explicitly says Inkling is a starting point, not a finished product.

**Executive Action Plan** If you're a technology or AI leader watching this space, here's what to do with this information. First, run a genuine build-versus-buy audit on your AI infrastructure. If your organization has proprietary data that could train a specialized model — internal documentation, historical decisions, domain expertise — price out what fine-tuning on an open-weight foundation would cost versus your current API spend over 24 months.

The math is shifting, and Inkling is one more data point that open-weight options are becoming enterprise-grade. Second, if you have ML capability in-house, download Inkling and benchmark it on your actual use cases. Don't rely on general leaderboards.

Mollick's critique is fair as a general statement, but your specific domain might be one where Inkling performs well or where fine-tuning rapidly closes the gap. Third, watch the Tinker platform closely. The model is the door; Tinker is where Thinking Machines makes money.

If their fine-tuning tooling is genuinely accessible and the results hold up, it could represent a meaningful shift in how mid-market companies approach AI deployment — less reliance on frontier API providers, more ownership of the models powering their core workflows.