Thinking Machines releases Inkling 2 Small: 594 billion parameters, but only 17 billion active

If you have ever tried to choose a model for agents, you know the dilemma: capable models are expensive and closed, while open models are cheaper but weaker on long, multistep tasks. Mira Murati's Thinking Machines is offering its own answer: Inkling 2 Small, a multimodal model whose weights are expected to be released under the Apache 2.0 license next week.

What's inside

Inkling 2 Small is built on a sparse Mixture-of-Experts architecture. The easiest way to picture it is as a large clinic: it has 384 doctors, or experts, but only the eight best suited to each patient, or token, are called in. The model has 48 decoder layers and 594 billion parameters in total. Only 17 billion are active during generation.

In practice, this means the speed and cost of each token are closer to those of a 17B-class model, while substantially more "knowledge" is stored in the weights. But there is a caveat that is often overlooked: all 594 billion parameters still have to be held in memory. Even at 8-bit precision, that amounts to hundreds of gigabytes, so the model will not run on a single gaming GPU. This is a job for servers or multi-GPU clusters.

The model also claims visual input, configurable reasoning depth and a context window of up to 1 million tokens. Reasoning depth means users can deliberately choose how long the model "thinks": less for simple requests and more for difficult ones. This provides a direct lever for controlling the inference bill.

Focus on agents

The developers position the model for multistep agent workflows and claim three capabilities. First, it can visually inspect its own output: for example, it can generate an interface, examine a screenshot and correct the layout. Second, it adapts to different agent harnesses, meaning it does not require a prompt and toolset tailored perfectly to the model. Third, it can orchestrate its own fine-tuning pipelines.

That last claim sounds impressive, but it should be independently verified. It comes from the developers, and the announcement provides no independent confirmation.

The numbers are as follows. The model scored 31 on the Artificial Analysis Intelligence Index. That is a modest result compared with heavyweight closed flagships, and it is fair to say that the model is not a leader in general intelligence. In DesignArena, however, a ranking for frontend and interface generation, it places third worldwide, behind only proprietary heavyweights. For teams building UI agents and interface generators, that metric is more relevant than the general index.

What to do with it

Inkling 2 Small is currently available for inference and fine-tuning through Tinker, Thinking Machines' own service. The full weights are due next week, after which users will be able to deploy the model themselves or fine-tune it on their own data without being constrained by API terms: Apache 2.0 permits commercial use.

The practical takeaways are:

• If you are building frontend or design agents, add the model to your benchmark as soon as the weights are released and compare it on your own tasks, not public rankings.
• If you do not have a cluster, calculate hosting costs in advance. The 17 billion active parameters reduce compute requirements, but not memory requirements.
• If you need a custom agent for a narrow domain, fine-tuning under an open license makes the model a strong candidate, especially if inference providers deploy it quickly.

Limitations

For now, everything rests on the lab's own announcement. There are no independent measurements of speed, cost or long-context behavior, while the score of 31 on the index suggests that no miracles should be expected in general reasoning. The weights have also not yet been published, so the Apache 2.0 promise should be considered fulfilled only when the files appear.

Even so, the development itself matters: one of the industry's most prominent labs is entering the open-model market with a large multimodal MoE and a permissive license. That gives the open agent-model market another serious contender, which means more choice and greater pressure on prices.