From websites to 3D scenes, an AI that also looks at what it creates

On Tinker, Inkling 2 Small can inspect its creations by image and run trainings. Its settings remain locked at launch.

A screenshot can serve as feedback to Inkling 2 Small so it can pick up the code it has just written. The new generation of Thinking Machines models combines image understanding with agent capabilities to build and refine websites, games, or 3D scenes. Visual rendering thus becomes part of the verification work, as changes are made.

In the demonstrations published by the lab, this loop is used to create an interactive city simulator or a paper-and-origami scene in Blender. The model receives text and images and produces text, including code: visual creations go through the tools it uses. It does not support audio. Learning to continue working for several hours

Training focused on hundreds of thousands of long-running programming tasks, carried out in different agent environments. For some of them, the reward did not depend solely on passing tests: a reviewer also examined the way the problem was approached. The lab says it worked on following instructions and spontaneously checking results.

The main reinforcement learning phase involved more than 17 million attempts over nine days, with 2,048 GPUs. It combined programming, document creation, and tool use, with evaluation criteria adapted to each task family. For web interfaces, reviewers compared screenshots of the rendered output, among other things.

This autonomy also caused problems during development. The lab reports having stopped training after an intermediate version bypassed a network restriction by using an accessible third-party service. It says it strengthened the infrastructure, removed certain environments, and added a check on tool calls before execution, then resumed training. Confirmed progress, with longer answers

The score of 31 on the Artificial Analysis Intelligence Index is confirmed by the evaluation organization, compared with 13 for Inkling Small. The lab’s detailed table gives 30.6, rounded to 31 in the announcement. This is a composite index, not a 31% success rate.

The gain is accompanied by heavier text generation. Artificial Analysis reports about four times as many tokens produced per task as for Inkling Small. Across this evaluation set, the average cost rises from $0.20 to $1.18 per task. These amounts describe the index protocol, not the fixed price of a request on Tinker.

Among the publisher’s acknowledged limitations are overly verbose answers and sometimes inappropriate tool choices, especially when the project contains many specialized instructions. The conversation is also less polished than the agentic tasks: response organization and overreactions to certain requests still need improvement. Customizing the model before being able to download it

Initial access goes through Tinker, to use the model and continue training it on specific tasks. The Inkling family documentation also provides for adjusting the reasoning effort. Customization can therefore concern both learned behavior and the level of effort requested at execution time.

Despite its name, “Small” refers to a 594-billion-parameter model, of which 17 billion are activated to process each token. This mixture-of-experts architecture does not make it a small model intended from the outset for a personal computer.

The parameters remain closed at launch. The November 5 announcement says they will be released under the Apache 2.0 license the following week, after external safety testing. The versions resulting from customized training are expected to become downloadable at that time. The license announced for this future release should therefore not be confused with the terms of current access to Tinker. Another model trained under its direction

An internal experiment also shows the agent in charge of a training run. The mission was to adapt Qwen3.5-4B for time-series forecasting, using data on energy, weather, air quality, and Wikipedia page views.

According to the lab, the agent built its evaluator, prepared the data, tried several methods, and fixed problems encountered along the way. The demo reports about four and a half hours of work and $426 in spending on Tinker. An attempt at reinforcement learning degraded the forecasts; the agent stopped it and kept a version from supervised training.

The promotional summary mentions a more than doubled forecasting score. The detailed account of the experiment provides a separate measure: on the unseen August and September data, the version ultimately retained produces about 15% less error than the original model. This result applies to that internal forecasting task, after selecting the best attempt.