Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026. The model is available through Tinker for sampling and fine-tuning, giving developers a new option for work that spans code, visual feedback and multiple steps.
What the launch makes available
The company describes a mixture-of-experts model with 594 billion total parameters, 17 billion active parameters, image inputs and adjustable effort. Its announcement says initial access uses closed weights on Tinker. The lab intends to release the full weights under Apache 2.0 after external safety testing and red-teaming, so a plan to self-host should wait for the actual release artifacts and terms. Launch announcement
The accompanying model card describes the model’s capabilities, evaluations and limitations. Those materials are a starting point for choosing a test workload. They do not establish the quality or reliability of a particular customer deployment. Model card
Separate model selection from serving decisions
The distinction between hosted access and downloadable weights affects procurement, engineering and data governance in different ways. A team can investigate behavior through a service while separately checking whether its intended data is permitted there. A future self-hosted deployment would bring additional questions about supported runtimes, hardware, maintenance and responsibility for upgrades.
For an engineering evaluation, pick a task that needs a coherent sequence rather than one impressive answer. A useful example is repairing a small web application against a written acceptance checklist. The model should inspect the project, change the appropriate files, execute relevant checks and explain the result. Each stage provides evidence about a different failure mode.
Include a visual requirement that can be checked objectively, such as keeping a navigation control visible at a narrow viewport or placing a label beside the correct field. This makes the model’s image-input capability relevant to the test. Save the before and after screenshots, the command outputs and the final diff so reviewers can judge the result without relying on the model’s description.
Measure the whole task at the chosen effort
Adjustable effort should be treated as an experimental setting. A higher setting may improve a difficult result while increasing response time or generated work. Compare configurations using the same task set and record completion, review burden and total spend. A fast first answer followed by several repair rounds can be more expensive than a slower successful run.
The model’s reported parameter count also does not directly determine the cost of a workflow. The serving provider, context size, number of turns, retries and tools all contribute. Track those quantities alongside quality so the pilot produces a decision the operations team can use.
Inkling 2 Small adds another model to evaluate for extended technical work. Teams interested in its planned open-weight path can prepare a reproducible task set now, then repeat that set when the downloadable release and its supported inference stack are available.