Thinking Machines debuts Inkling Small open source AI model nearing performance of predecessor at about 1/4…

Thinking Machines has released Inkling Small, an open source AI language model, only two weeks after shipping Inkling, its first open source language model. The new model is positioned as roughly one-quarter the size of its predecessor while approaching that model’s performance. The company is a well-funded startup led by Mira, former chief technology officer at OpenAI.

On product mechanics, the headline claim is size versus capability: Inkling Small aims to stay close to Inkling’s results at about one-fourth the footprint. That framing treats the original Inkling release as the performance reference and the smaller model as a compression of the same line rather than a separate product family. Both releases are open source language models from the same vendor, so the comparison is within one stack, not across unrelated systems.

For engineers and builders, a model that nears prior performance at about a quarter of the size changes deployment math: lower memory and compute budgets for serving, easier local or edge trials, and cheaper iteration when you are evaluating open weights. The two-week gap between Inkling and Inkling Small also signals a fast open-source release cadence, so teams tracking this line should treat the first model as a baseline that may be superseded quickly rather than as a long-lived default.

In market terms, a well-funded startup led by a former OpenAI CTO shipping open source language models is competing on openness and iteration speed as much as on raw capability. Following a first open model almost immediately with a much smaller variant that still aims near the predecessor’s performance is a classic open-weights play: give the community something usable under tighter resource constraints without fully abandoning the larger model’s quality bar.

What to watch next is whether Inkling Small’s “nearing” claim holds on the workloads builders actually care about, and how the company documents the size-to-performance tradeoff relative to full Inkling. The practical move is to treat Inkling as the reference and Inkling Small as the efficiency option from the same open source line, then re-check requirements when the next drop lands—especially given how quickly the second model followed the first.