Mira Murati, the architect behind ChatGPT's product strategy, just bet against the thing she helped build. On Wednesday, she released Inkling, an open-weight model from her new company Thinking Machines Lab. Unlike the walled frontier models, anyone can download it, modify it, and run it locally.
The move is not sentimental. It's a specific strategic claim: open systems with different training approaches can outcompete closed ones at particular tasks without ever matching their general-purpose scale.
Inkling is structured as a mixture-of-experts with 975 billion total parameters but only activates roughly 41 billion per task. It trained on 45 trillion tokens of text, image, audio, and video, a multimodal dataset, though for now it outputs only text. The numbers position it in the middle of the frontier field: bigger than Sonnet 5, smaller than Opus 4.8.
What matters is not the scale but the thesis it tests. Murati is betting that the frontier lab playbook has become predictable: chase the largest possible context window, the densest parameter count, the broadest capability footprint. Win on general benchmarks, ship behind paywalls, extract value from lock-in. Her counter is that specialized open systems, trained on different data and released under permissive licenses, can own specific tasks and workflows where enterprises need ownership and control.
This is not a new argument. Every open-source champion makes it. But it matters when someone who shipped ChatGPT makes it. Murati has the credibility to be wrong at the strategic level and still move markets; she's not some scrappy startup founder telling you the incumbents are overconfident. She's an insider who watched the playbook work and decided to build against it.
The model arrived with no performance claims, no benchmarks, no leaked internal evals suggesting it competes with Opus or Sol. That absence is itself the argument: Thinking Machines is not chasing the benchmark leader board. The company spent eighteen months building infrastructure "largely out of public view," per their release, suggesting a different optimization target than the race-for-SOTA the industry has become.
Two details strike me. First, the timing. Gemini 3.5 Pro drops Thursday with its 2-million-token context window. GPT-5.6 Sol just went live. The frontier field is contracting toward raw capability density. Murati is releasing an open model the same week, which says something about her irrelevance by that metric or her willingness to play a different game entirely. Second, the architecture. Mixture-of-experts models only activate a fraction of their parameters per forward pass. They're cheaper to run, yes, but they're also harder to understand and reason about. Open-sourcing a MoE model is almost a dare: figure out how this actually works. Build on top of it. The frontier labs lock capability in closed systems. Murati locked complexity into an open one.
She spent a year and a half saying nothing. In her silence, she was either learning the frontier labs' playbook was brittle or learning it was not. Wednesday's release suggests the former. The verdict will arrive in the developer graphs, whether open-weight systems start carrying workloads the paywall models do not serve well. Not whether Inkling beats Opus on eval leaderboards. Nobody at Thinking Machines is claiming that. The question is whether open systems owned by developers matter more in 2026 than they seemed to in 2025. Murati is betting yes.
Top comments (0)