Mira Murati's Thinking Machines Releases Its First Open-Weight AI Model, Inkling

Mira Murati's AI startup finally shipped something you can actually touch.

Thinking Machines Lab released Inkling Wednesday, its first public model since Murati left her post as OpenAI's chief technology officer to start the company roughly eighteen months ago, according to TechCrunch. Unlike ChatGPT, Claude, or Gemini, Inkling is open-weight. Anyone can download it, inspect it, and rebuild it for their own purposes.

What It Actually Is

Inkling is a mixture-of-experts system with 975 billion total parameters, but it only activates about 41 billion of those for any single task, according to both TechCrunch and VentureBeat. That design keeps a massive model cheaper and faster to run than its full size implies.

It trained on 45 trillion tokens spanning text, images, audio, and video, and TNW reports it can reason across all four formats natively. For now, though, outputs are text-only, including code and structured data. A smaller companion, Inkling-Small, was previewed alongside it, with VentureBeat citing 276 billion total parameters and TNW citing 12 billion active parameters for that variant.

It runs under an Apache 2.0 license, VentureBeat noted, which is about as enterprise-friendly as open licensing gets. No murky terms, no restrictions that make legal departments nervous.

Not Claiming to Be the Best

Thinking Machines said flatly in its own release materials that Inkling is "not the strongest overall model available today, open or closed," according to TechCrunch. That's an unusual thing for a company to admit about its own product.

The benchmarks back it up. VentureBeat reported that China's GLM 5.2 beats Inkling badly on hard reasoning and coding tasks, putting up 82.7 on Terminal Bench 2.1 against Inkling's 63.8, and outscoring it on SWE-Bench Pro, 62.1% to 54.3%.

Where Inkling does hold its own: VentureBeat reported it hit 77.6% on SWE-bench Verified, beating Nvidia's Nemotron 3 at 71.9%. TechCrunch reported Thinking Machines' own claim that Inkling matches Nemotron 3 Ultra's coding performance while using a third as many tokens. On voice understanding, VentureBeat put Inkling at 91.4% on VoiceBench, trailing Google's Gemini 3.1 Pro at 94.4%.

So it's a solid mid-tier generalist, not a champion. Thinking Machines seems fine with that.

The Real Bet: Customization Over Domination

The company's actual argument, laid out in a manifesto published the week before the release, is that AI "trained centrally by one company and then set in stone" loses to AI that organizations shape themselves, according to TechCrunch. Customers are expected to fine-tune Inkling through Thinking Machines' own platform, called Tinker, rather than use it out of the box.

That shifts real responsibility onto the customer. If a company fine-tunes Inkling and the result behaves badly or unsafely, that's on the company doing the tuning, not on Thinking Machines. TNW reported this explicitly: customers who fine-tune Inkling "own the result" and "carry the safety risk" of what they build. Fine-tuning also isn't a task for amateurs. It requires real machine learning talent, which limits who can actually take advantage of the open weights beyond running the model as-is.

Thinking Machines pointed to a project with hedge fund Bridgewater as proof the strategy works, claiming a model fine-tuned on Bridgewater's financial expertise scored 84.7% on financial reasoning benchmarks, beating proprietary models at lower cost, according to TNW. That number came from the two companies' own joint evaluation, not an independent test, so it deserves the same skepticism as any vendor benchmark.

VentureBeat's coverage flagged something the other outlets underplayed: Thinking Machines specifically designed Inkling to answer directly on subjects that might otherwise get censored, framing it as a more trustworthy option for enterprises that want factual answers regardless of how sensitive the topic is. That's a notable selling point in a market where users have grown wary of chatbots refusing to answer straightforward questions.

Fast, and Built on Borrowed Parts

Murati's team says it built Inkling in about nine months, compared to the roughly three to five years OpenAI and Anthropic took to ship their first major products, a comparison TechCrunch attributed to Thinking Machines itself. Part of how they moved that fast: TNW reported the company used distillation, training Inkling partly off other open models including Moonshot's Kimi K2.5, rather than building everything from scratch.

Databricks announced it's a "day zero" launch partner, making Inkling available through its Unity AI Gateway so enterprise customers can plug it into coding agents like Cursor and OpenCode while keeping data inside their own governed environment, according to the company's own blog post.

The open question now is adoption. Plenty of enterprises say they want control and customization over raw benchmark scores. Whether they'll actually put in the engineering work to fine-tune a 975-billion-parameter model, rather than just paying for the best closed model available, is the bet Thinking Machines is making with its entire business.