Thinking Machines, the AI lab from former OpenAI CTO Mira Murati, has released Inkling Small. The model arrives with a clear message: a smaller system can still compete closely with a larger one when efficiency, reasoning performance, and fine-tuning access are treated as core design goals.
According to Artificial Analysis, Inkling Small scores 40 on the Intelligence Index. That puts it just one point below Inkling, which scores 41, even though Inkling Small has less than a third of the parameters: 276 billion total and 12 billion active.
A smaller model with a close score
The central comparison is straightforward. Inkling Small does not surpass Inkling on the overall Intelligence Index, but it comes very close while using a much smaller parameter profile. Artificial Analysis says no open model of equal or smaller size scores higher.
That matters because model size is often treated as a shorthand for capability. The numbers here complicate that view. Inkling Small is presented as an open-weights reasoning model that gives up very little on the headline Intelligence Index score while reducing the amount of model capacity active during use.
The gap between 40 and 41 is narrow. The source does not say that Inkling Small is broadly better than Inkling, and the distinction is important. What it does show is that the smaller model can sit near its larger sibling in a widely cited performance comparison while occupying a different efficiency profile.
Where Inkling Small beats Inkling
Inkling Small also outperforms Inkling on several coding and reasoning tests. The source names Humanity's Last Exam and GPQA Diamond as examples, with Inkling Small scoring 32% versus 30% on Humanity's Last Exam and 89% versus 87% on GPQA Diamond.
Those results make the release more than a simple smaller-model variant. In some measured areas, the compact model is ahead of the larger one. For developers and teams evaluating reasoning models, that kind of uneven profile is important: the best choice may depend on the task rather than the model family hierarchy.
At the same time, Inkling Small is not stronger everywhere. It falls behind on agent-based tasks and factual knowledge. That creates a more nuanced picture: the model appears especially interesting where coding, reasoning, and token efficiency are high priorities, but it may not be the best fit for every workload described in the source.
Token efficiency is the clearest differentiator
The most practical difference may be how much output the model uses per task. Inkling Small averages 24K output tokens per task. The source compares that with 45K for Deepseek V4 Flash and 78K for GPT-5.4 mini.
That comparison points to a simple operational implication: a model that solves tasks with fewer output tokens can be easier to work with in settings where long generations are a constraint. The source does not provide cost data, latency data, or deployment benchmarks, so those points should not be inferred directly. But output-token efficiency is still a concrete measure of how compact the model's responses are during tasks.
For users building with reasoning models, token use can shape the experience. Long outputs may contain more intermediate material, but they can also be harder to review and integrate. A lower average output token count suggests that Thinking Machines is emphasizing concise reasoning behavior alongside model capability.
Open weights, multimodal input, and fine-tuning
Inkling Small handles text, image, and speech inputs. It also has a 256K-token context window, which gives it room to process large amounts of input in a single context.
The model ships under Apache 2.0, and its weights are on Hugging Face. Users can also fine-tune it in the browser via Tinker Playground. These details matter because Thinking Machines is positioning its models as a foundation for fine-tuning with users' own data.
That positioning connects the technical release to a broader product idea. Rather than offering only a fixed model, Thinking Machines is presenting Inkling Small as something users can adapt. The source notes that some see this as the next frontier in AI.
What the release signals
Inkling Small is not described as the largest model in the comparison, and that is the point. Its strongest claim is not raw scale. It is the combination of near-Inkling Intelligence Index performance, specific wins on coding and reasoning tests, lower average output tokens, multimodal input support, a long context window, open weights, and browser-based fine-tuning.
The release also underlines a shift in how AI models may be judged. Size still matters, but it is not the only axis. A model can be evaluated by how close it gets to larger systems, where it beats them, where it falls short, how many tokens it needs, and how easily users can adapt it to their own data.
Based on the source, Inkling Small is best understood as a bet on efficient, customizable reasoning rather than a simple race toward more parameters. For Thinking Machines, that makes the model a statement about direction as much as performance.