On Tuesday, Mira Murati released her first AI model since leaving OpenAI.
If you don't know her name: Murati was Chief Technology Officer at OpenAI from 2022 to 2024. She was the person responsible for shipping ChatGPT, GPT-4, DALL-E, and Sora. She knows what frontier AI looks like from the inside more than almost anyone. She left OpenAI in late 2024, raised $2 billion to start Thinking Machines Lab, and spent 18 months building something. This week she published it.
It's called Inkling. It's 975 billion parameters. It's released under Apache 2.0 — no restrictions, full commercial use, modify and deploy however you want. And it's explicitly not designed to be the most powerful model available.
That last part is the thing worth paying attention to.
Built to be replaced by yourself
Most AI releases position themselves as the strongest model available. Use us. We're the best. Inkling's announcement says something different: "Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization."
That's a fundamentally different play.
The premise is that you don't actually want the world's most capable general model. You want a model that's good at your work. Your documents. Your terminology. Your cases. Your policies. A general model tries to know everything and ends up knowing your specific domain imperfectly. A fine-tuned model trained on your data can outperform far larger models on the tasks you actually care about.
Murati built Inkling to be that starting point.
The architecture supports it. It's a Mixture-of-Experts model with 975 billion total parameters, but only 41 billion are active for any given token — inference is cheap relative to what the parameter count implies. It has a 1-million-token context window. It reasons natively over text, images, and audio. And it's wired directly into Tinker, Thinking Machines' fine-tuning platform, where you can upload your own dataset and train a custom version today.
The base model isn't sandbagging. It scored 87.9% on GPQA Diamond (graduate-level reasoning), 77.6% on SWE-bench Verified (real software engineering tasks), and 97.1% on AIME 2026 (competition math). These are frontier numbers, not "good for open source" numbers. You're getting a serious model as your starting point.
The censorship angle that matters for NGOs
VentureBeat's coverage led with two things: low cost and resistance to censorship. The second one is worth unpacking.
Thinking Machines trained Inkling to answer directly on topics that mainstream models often refuse or hedge on — political content, conflict zones, morally complex questions, legally sensitive material. On the FORTRESS benchmark, which specifically measures whether a model engages with content that more constrained models dodge, Inkling posted the highest adversarial-safety score among open-weight models at 78.0%.
For most small businesses this won't be relevant. But for a meaningful slice of the orgs we work with — NGOs doing advocacy, teams writing policy on sensitive topics, nonprofits in harm reduction or conflict response — mainstream AI models have a real tendency to be unhelpfully cautious. The refusals feel arbitrary because they are: they're set by a vendor optimizing for liability across a billion users, not for the legitimate needs of your specific work.
When you fine-tune Inkling on your organization's data, you get that directness plus your own content policies. You're not inheriting whatever OpenAI or Google decided is acceptable for their entire user base.
The irony nobody's talking about
There's a subplot here worth noting.
The U.S. government spent significant political capital this year restricting access to Chinese AI models — partly over concerns that Chinese labs were using "distillation" techniques: training models on outputs from OpenAI and Anthropic's systems to capture their capabilities without the research and development cost.
Thinking Machines' own materials acknowledge that Inkling's post-training data included outputs from Kimi K2.5, an open-weight model from Chinese startup Moonshot AI. They say the next model will use fully self-contained training instead.
So: a U.S.-based startup founded by a former OpenAI executive used Chinese model data to train an open-weight model that any organization anywhere can now download and customize. The clean national-security narrative around AI is messier in practice than the headlines suggest. The technology is moving faster than the politics, and the boundaries are porous in every direction.
What to do with this right now
If you're using AI for internal knowledge work: Inkling is worth testing as a base. A 1-million-token context means you can pass an entire document library into a single session. The API is live via Tinker and Together AI at around $1.87 per million input tokens.
If generic AI keeps failing on your specific domain: Seriously consider fine-tuning. This is no longer an enterprise-only option. Feed Inkling your case files, program documentation, historical reports. The model you get back will outperform a general-purpose model on your actual work, not some benchmark you don't care about.
If mainstream AI tools refuse too much for your work: Test Inkling before building elaborate prompt workarounds. It was built to engage directly. That might be the simpler fix.
If you have data residency requirements: The weights are on Hugging Face under Apache 2.0. You can run this on your own infrastructure — a requirement for some public sector deployments where cloud APIs aren't an option.
The broader pattern: the people who built closed AI are increasingly going indie and publishing their work openly. The tools that were locked inside a $100 billion company last year are now downloadable, modifiable, and deployable by a 12-person team.
The question isn't whether small orgs can access frontier-grade AI anymore. It's which ones will use that to actually change how they work before their peers figure it out.
We help teams work out where fine-tuning makes sense for their specific situation, and we run the setup when it does. Usually a sprint, not a strategy.