Thinking Machines Releases Inkling 2 Small, a 594B MoE Model

Thinking Machines Lab has released Inkling 2 Small, a 594B-parameter mixture-of-experts model with 17B active parameters, according to MarkTechPost. The release matters because its total size, active-parameter count and index score give developers and researchers a concrete new point of comparison when they are choosing, testing or fine-tuning models for practical work.

The name itself carries useful information. Inkling 2 Small is presented as a smaller member of a family or generation, rather than as the largest possible system. That positioning is common in model releases: a smaller variant is often intended to be easier to run, cheaper to serve or simpler to adapt than a flagship version, even when its headline parameter count still sounds very large. Readers should therefore resist judging the release by the word Small alone, or by the 594B figure alone, and instead look at how the two numbers work together.

What was released?

According to MarkTechPost, Thinking Machines Lab released Inkling 2 Small as a 594B-parameter mixture-of-experts, or MoE, model with 17B active parameters. The report also states that the model scores 31 on the Artificial Analysis Intelligence Index, that it runs on Tinker, and that Apache 2.0 weights are planned.

Those are the central facts available in the report, and they frame the questions that follow. What kind of architecture is being described? How should a reader interpret a score of 31 without more context? What does it mean for a model to run on a particular platform, and what would planned open weights change for potential users? The answers to those questions determine whether Inkling 2 Small is mainly a research announcement, a platform offering, a future open-weight release, or some combination of the three.

It is also worth noting what the report does not claim. It does not provide, in the summary available to Grafy News, a detailed breakdown of benchmark categories, pricing, context length, training data, or specific applications. A responsible reading therefore treats the release as an initial signal rather than a full evaluation. The most useful response for a practitioner is not to accept or dismiss the headline numbers at face value, but to ask how the model performs on the tasks that matter for their own work, and under what conditions it can actually be accessed.

What do 594B total and 17B active parameters mean?

In a conventional dense model, the total parameter count is a rough guide to the size of the network and, in many cases, to the amount of computation involved in generating a response. A mixture-of-experts model complicates that simple picture. In an MoE design, the model contains many specialised sub-networks, often called experts, and a routing mechanism selects only a subset of them for a particular token or input. The result is that the model can hold a very large total number of parameters while using only a fraction of them at any one moment.

That is the distinction between the 594B total parameters and the 17B active parameters reported for Inkling 2 Small. The larger figure describes the overall capacity stored in the model. The smaller figure describes the portion activated during computation. For readers comparing models, both numbers matter, but for different reasons. Total parameters can affect storage requirements, memory demands and the resources needed to load the model. Active parameters are more closely related to the computation performed for a given generation, and therefore to questions about speed and serving cost, although hardware, implementation, batching and other factors also influence real-world performance.

The ratio between the two figures is striking. Seventeen billion active parameters out of 594 billion total is a small fraction, which illustrates the MoE idea in concrete terms: most of the model's capacity is available across different inputs, but only part of it is used for any single computation. That does not automatically make the model efficient in every deployment, nor does it guarantee a particular level of quality. It does, however, explain why a model can be described simultaneously as very large and as comparatively economical to run, depending on which figure is being emphasised.

For developers, the practical lesson is to compare like with like. A dense model with a given parameter count and an MoE model with the same total count may behave very differently in cost and latency. Conversely, two MoE models with similar active counts may differ in total capacity, routing design and quality. Headline parameter counts are therefore a starting point for a conversation, not a substitute for task-based testing.

What does a score of 31 on the Artificial Analysis Intelligence Index tell us?

According to MarkTechPost, Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index. An index score of this kind is intended to summarise performance across a range of tasks into a single comparable number. That makes it useful for a first scan of the landscape: a reader can place a new model alongside others that have been measured on the same scale and get a sense of its general position.

A single score also has clear limits. It cannot show whether a model is strong at coding but weaker at reasoning, or good at summarising but unreliable at following detailed instructions. It cannot reveal how the model handles long documents, tool use, multilingual input, safety constraints or the specific formats used in a particular product. Two models with similar index scores may feel very different in use, and a model with a lower score may still be the better choice for a narrow task where its strengths align with the requirement.

The most productive way to use the figure is as a prompt for further questions. Which tasks contribute to the index? How do competing models score on the same index? Are the differences between scores large enough to matter in practice, or are they within a range where other factors such as price, latency, licensing and platform support become decisive? Those questions turn a headline number into a tool for decision-making.

Readers should also be careful about treating any index as neutral or final. Benchmarks reflect choices about what to measure and how to weight different abilities. A model can be optimised for the kinds of tasks a benchmark emphasises, and real-world work often involves messy, underspecified problems that are harder to capture in a test. The score of 31 is therefore best understood as one piece of evidence, to be combined with hands-on evaluation on representative examples from the user's own domain.

How does it run, and what does Apache 2.0 mean for access?

According to MarkTechPost, Inkling 2 Small runs on Tinker. That detail matters because a model is only useful to most people through some route of access: an application programming interface, a hosted service, a fine-tuning platform, downloadable weights, or a combination of these. Naming the platform tells readers where the model is currently available to run, and therefore where they would go to test it, integrate it, or adapt it.

The report adds that Apache 2.0 weights are planned. Apache 2.0 is a permissive open-source licence. If weights are released under that licence, users would generally be able to download, use, modify and redistribute them under the terms of the licence, subject to its conditions. That would be a significant change in access, because it would allow teams to run the model in their own infrastructure, subject to the hardware and engineering demands of a model of this scale, rather than relying solely on a hosted platform.

The word planned is important. A plan to release weights is not the same as weights that are available today. Readers should distinguish between what can be used now on Tinker and what may become available later under Apache 2.0. They should also consider the practical requirements of open weights: serving a model with 594B total parameters is a substantial infrastructure task, even with only 17B active parameters, and fine-tuning or evaluating it will require appropriate hardware, software and expertise. Open weights expand freedom, but they do not remove cost or complexity.

For many organisations, the relevant question will be which route fits their constraints. A hosted platform may be faster to adopt and easier to scale. Open weights may offer more control, privacy or customisation. The report, as summarised, points to both possibilities, but the timing and details of the planned release will determine how much weight to place on each option.

How could you try or compare it in Grafy?

Grafy does not claim to offer Inkling 2 Small as a specific feature, and readers should check the current model catalogue for the exact models available at any given time. What Grafy can offer is a practical way to explore the questions raised by this release, particularly around model comparison. Grafy Chat lets users chat with many AI models side by side, which is well suited to testing how different models handle the same prompt, instruction or coding problem. That side-by-side approach mirrors the lesson from the Inkling 2 Small numbers: total size and index scores are only a starting point, and direct comparison on real tasks provides stronger evidence.

A related Grafy Blog guide on asking several AI models at once and letting a judge pick the best, Council of AI Models, describes a workflow that can be applied to any new release. Instead of relying on a single headline score, a reader can run the same set of representative prompts across multiple models, compare the outputs, and judge which model best fits a particular use case. For a release like Inkling 2 Small, that could mean testing reasoning, writing, coding or summarisation tasks, depending on what the reader actually needs to do.

Grafy Reader offers a complementary route for research-oriented readers. Reader is designed to read, search and question PDFs and papers with AI, which can help users examine source material, technical reports or evaluation documents more closely. When a new model is announced, the ability to search a document for specific claims, compare sections, and ask targeted questions can be more valuable than a summary alone. Readers who want to understand the evidence behind a release, rather than just its headline figures, will find that approach useful.

The Grafy News Desk perspective is straightforward: treat Inkling 2 Small as a prompt to test, not just to read about. Identify a small set of tasks that reflect real work, run them across several models, and record where each model succeeds or fails. That method works whether the model of interest is available through a platform such as Tinker, through future open weights, or through a multi-model workspace where comparison is built in.

Frequently asked questions

How many parameters does Inkling 2 Small have in total?

According to MarkTechPost, Inkling 2 Small has 594B parameters in total. That figure describes the model's overall parameter count, which is distinct from the 17B active parameters reported for a given computation. The distinction matters when comparing efficiency claims, because total size and active size answer different questions about a mixture-of-experts design.

What score did Inkling 2 Small achieve on the Artificial Analysis Intelligence Index?

According to MarkTechPost, Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index. That single number is best treated as a summary signal rather than a complete verdict on usefulness. Readers should ask which tasks the index covers, how other models score on the same scale, and whether those tasks match their own work before drawing practical conclusions.

On what platform does Inkling 2 Small run?

According to MarkTechPost, Inkling 2 Small runs on Tinker. The report also says Apache 2.0 weights are planned, which concerns future access to the weights rather than the current running platform. Readers interested in fine-tuning or deployment should therefore distinguish the platform named in the report from the licensing plan, and check the original announcement for the latest availability details.

Source: MarkTechPost

---

Source: MarkTechPost — Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active