“We believe in keeping the weirdness alive.” It’s not the usual sterile press release, but the manifesto that accompanied the release of Inkling by Mira Murati’s lab, the former OpenAI chief technology officer. Thinking Machines Lab enters the public stage with an open-weight model—anyone can download it—and a disarming admission: it’s not the best. In a race where every percentage point on benchmarks is heralded as a revolution, upfront imperfection feels less like a marketing strategy and more like a political statement.

The move shouldn’t be dismissed as eccentricity. It comes from someone who, inside OpenAI, saw firsthand what it means to box research into metrics that reward ever larger and more opaque models, accessible only via APIs. Releasing an LLM with open weights, even if not state-of-the-art, shifts the field: it gives developers and companies back the ability to operate self-hosted, without sending data to third parties, without relying on a subscription fee and an endpoint whose policies can change overnight. For many sectors—healthcare, finance, government, manufacturing—‘good enough’ running on-premise, under one’s own control, is worth more than the perfect model living on someone else’s servers.

There’s a deeper thesis in this move, and it concerns the very structure of the industry. Large centralized labs push toward a future where artificial intelligence is a cloud service, metered, with no transparency on training and zero room for radical customization. Inkling, with its technical ordinariness, is an act of resistance: a reminder that the ecosystem needs open building blocks that can be taken in-house, fine-tuned on proprietary data, quantized to run on owned hardware. It’s no accident that the model ships with a license allowing download and use without onerous commercial restrictions. The goal isn’t to climb leaderboards, but to lower the barrier for those who want to govern their own inference stack.

Who wins? Primarily actors operating under data residency regulations like Europe’s GDPR, or simply under industrial secrecy requirements. Being able to download an LLM and run it on a local cluster, even without cutting-edge GPUs, can mark the difference between using AI and sitting on the sidelines. Of course, compute resources are needed—and here the vagueness of Inkling’s tech sheet (no details on VRAM, precision, minimum requirements) is a limitation. But the signal is clear: not everything that matters needs to sit behind an API endpoint’s glass case.

Murati’s move also speaks to hardware vendors and system integrators. If open models stop being hobbyist afterthoughts and become production assets, demand grows for servers and workstations optimized for local inference. It’s not just about top-tier graphic cards: the market is already exploring powerful CPUs, dedicated accelerators, and large RAM pools to push self-hosting without necessarily competing on enterprise GPU pricing. Inkling can be read as an invitation to take this path seriously, separating the AGI race from the daily need to automate processes, analyze documents, and generate reports.

One open question remains: how hard is it, in practice, to put it into production? Without official benchmarks on real-world performance—tokens per second, latency, energy consumption—the model is currently more a political promise than an engineering tool. But it’s precisely the nature of the promise that’s interesting: it doesn’t sell itself as the best, it offers itself as the freest.