Thinking Machines’ Inkling 2 Mini puts more computation inside each generated token by repeating part of the model. The November 20 release combines that architectural choice with open weights and a local-device demonstration. The interesting tradeoff is how internal reasoning, memory requirements and observable behavior change together, rather than whether a smaller model can produce an impressive isolated example.

What is released

The Thinking Machines announcement describes a mixture-of-experts model with 26.4 billion total parameters and 3.4 billion active parameters. Accounting for its repeated passes, it reports 5.4 billion weight activations per token. The released model runs its middle layers three times. A separate variable-loop research variant supports experiments with different numbers of passes; results from that variant should retain their own label.

The company reports a roughly 20 GB memory footprint and demonstrates offline work on a MacBook Pro. Its comparison chart also uses a theoretical decoding ceiling at a specified memory bandwidth. That ceiling is not a measured laptop throughput result. Deployment decisions need measurements for the actual precision, runtime, context and device rather than a speed inferred from the chart.

More work can produce fewer visible tokens

Looping reuses parameters while increasing computation. That can be attractive when a system benefits from additional reasoning without storing a much larger model. But parameter count, activated weights and time to complete a task describe different costs. A comparison should make all three visible so an apparent reduction in output length is not mistaken for an equivalent reduction in total work.

The release also distinguishes reasoning gains from stored factual knowledge. Repeated passes do not automatically add facts to the weights. A local agent may still need retrieval or external tools, and those dependencies change the meaning of an offline capability claim. The relevant test should state which tasks were completed entirely locally and which required outside information.

For coding, a useful evaluation would include the surrounding agent harness, tools and verification process. A model can produce a shorter solution while spending more time in tool calls or requiring another repair attempt. Completed, verified tasks per unit of time and energy would help connect architectural efficiency with the user’s actual experience.

The same architecture makes oversight more complicated. Less verbalized reasoning does not establish that a model is doing less reasoning. Thinking Machines reports monitorability experiments and identifies the limits of those tests. They should be treated as evidence about specific evaluated settings, rather than a general assurance that internal computation remains fully understandable.

Evaluate the complete local system

A meaningful local comparison would record device configuration, runtime, quantization, context length, task success and sustained resource use. Thermal behavior and long-session stability can matter as much as a short speed test. These are proposed assessment dimensions, not additional measurements performed for this article.

Researchers should also separate the released model from experimental variants when reporting monitorability results. Changing the loop count, training or monitoring model can change the conditions being assessed. An open research artifact is valuable because others can investigate those conditions, not because one set of experiments settles the issue for every future deployment.

The release establishes a concrete model and an accessible architectural experiment. Its next persuasive evidence will come from reproducible task-level comparisons on common local devices, accompanied by careful reporting of what the model does internally and what a supervisor can actually observe.

Sources and media credits
  1. thinkingmachines.ai — source account

Thinking Machines’ Inkling 2 Mini release artwork. Supplier illustration, not an independently measured performance result. Image: Thinking Machines

Share this article

Post to XShare on LinkedInShare on WhatsAppEmail