Thinking Machines Lab's First Open Model, Inkling, Lands Day-0 on Databricks
Source: Databricks Blog — 2026-07-15
Summary
Databricks announced it's a day-0 launch partner for Inkling, the first open-weights model from Thinking Machines Lab: a 974-billion-parameter, 41-billion-active mixture-of-experts model released under Apache 2.0. The model is available on Databricks governed through Unity AI Gateway, meaning enterprises can adopt it with the same governance, access control, and monitoring layer they already use for other models rather than bolting on a separate integration path.
Key Takeaways
- Inkling is a 974B-parameter mixture-of-experts model with only 41B parameters active per inference pass — the MoE design routes each request through a fraction of the full model's weights, keeping inference cost closer to a much smaller dense model despite the model's total size.
- It ships under Apache 2.0, a genuinely permissive open-source license, unlike many "open-weight" releases that carry usage restrictions.
- Databricks integrated it through Unity AI Gateway from day zero, so it inherits existing governance (access control, auditing, monitoring) rather than requiring a separate bespoke integration for enterprises already standardized on that layer.
- This is Thinking Machines Lab's first open model release, making the choice of governed, enterprise-ready day-0 distribution (rather than a raw weights drop) a notable strategic signal about how the lab wants its models adopted.
Reel Script
Hook (~15s): A trillion-parameter-class model just went open source — but the more interesting number isn't 974 billion, it's 41 billion. That's the part of the model that actually runs on any given request.
Core Concept (~85s): Inkling is what's called a mixture-of-experts model, or MoE — instead of one giant neural network that processes every request with all of its parameters, it's built from many smaller "expert" sub-networks, and a routing mechanism decides which experts are relevant to a given input and only activates those. Think of it like a hospital with 974 billion dollars' worth of specialist equipment installed across every department, but any single patient visit only actually uses the roughly 41 billion dollars' worth of equipment relevant to their specific case — you get the benefit of enormous total capacity without paying the compute cost of running the whole hospital for every visit. That's why a 974-billion-parameter model can still be practical to serve: the "active" parameter count per request, 41 billion, is what actually determines inference cost and latency, not the total size.
Hands-On (~100s): The part worth diagramming is the deployment path Databricks built around this. Instead of Inkling landing as a raw weights download that every team has to figure out how to serve, govern, and monitor themselves, Databricks wired it directly into Unity AI Gateway on day zero — the same governance layer that already sits in front of other models on the platform. That means access control, usage auditing, and monitoring for Inkling come for free to any team already using Unity, rather than being a separate project. Combine that with the Apache 2.0 license — genuinely permissive, not one of the restricted "open-weight but not really open" licenses that have become common — and you get a model that's both technically efficient (MoE routing) and organizationally low-friction to adopt (governed day-0 access) at the same time.
Takeaway (~20s): If you've been holding off on trying a trillion-parameter-class model because of inference cost or governance overhead, Inkling's MoE efficiency plus its day-0 governed rollout on Databricks removes both excuses at once. Worth benchmarking against whatever you're currently running before your next model refresh.