Inkling 2 Small has 594 billion parameters, which gives the word "Small" considerable room for interpretation. Thinking Machines Lab's text-and-image model placed third in DesignArena's November 5 website-generation ranking, scoring 1340. Developers can already run and fine-tune it through Tinker, but its weights remain closed ahead of a planned release under Apache 2.0.
A strong result with a defined scope
Muse Wonder and Muse Spark 1.3 took the two positions above Inkling 2 Small. The top of the website-generation ranking looked like this:
| Position | Model | Score |
|---|---|---|
| 1 | Muse Wonder | 1366 |
| 2 | Muse Spark 1.3 | 1351 |
| 3 | Inkling 2 Small | 1340 |
The same ranking also includes GPT-6 Astra, Kimi K3 and Claude Opus 5.5. Their individual scores are not supplied in the report, so it does not support a detailed comparison with each of those models.
The category matters. Third place is a result for website generation, rather than evidence of the same standing across every programming task or general AI evaluation.
On the separate Artificial Analysis Intelligence Index, Inkling 2 Small scored 30.6, substantially above earlier Inkling versions. The report does not give those earlier scores, which leaves the size of that improvement unquantified.
A large model that activates only part of itself
The 594 billion-parameter total describes the model's full size. Its Mixture of Experts (MoE) architecture allows it to activate only a subset of those parameters for each request, reducing computation compared with running all of them at once.
That distinction helps explain how a model with such a large parameter count can limit the work performed on an individual request. It does not remove the resource demands of the complete set of weights, which remain substantial.
The report gives no active-parameter count. The total alone therefore cannot tell a developer how much of the model is used on a particular request.
Inkling 2 Small also supports adjustable reasoning depth. The available information does not specify the settings or quantify their effect on performance, speed or cost.
Coding with visual feedback
Inkling 2 Small can create websites, games and three-dimensional projects, then examine images of the results and make corrections. Its coding process can therefore include a feedback loop in which the generated result becomes input for the next revision.
For website and interface work, appearance is part of the output being judged. A model that can inspect the result has a way to respond to what it produced, rather than ending its work when the code has been written.
Visual feedback is a capability, however, rather than a guarantee that every revision will be an improvement. The reported ranking does not isolate how much this feature contributed to the model's score.
Training also included extended software-development tasks. The examples given were:
- Real repository work, exposing the model to tasks within existing software projects.
- GPU performance work, involving efforts to improve how graphics processors are used.
- Autonomous model training, in which it trained other models.
These examples describe the training program. No separate benchmark results are supplied for repository work, GPU performance or autonomous training, so the website-generation result should not be stretched into a claim about performance in all three.
Tinker access now, downloadable weights later
Developers currently have access through Tinker for running and fine-tuning Inkling 2 Small. Until its weights are published, access remains on Thinking Machines Lab's own infrastructure.
The company plans to release the weights under Apache 2.0 during the week following the November 6 report. That release is intended to follow external safety reviews and testing of the model's resistance to attempts to bypass its restrictions.
The release is still planned. The report gives no completed review results or exact publication date. Access through Tinker is available now, while the weight release remains a separate, future step.
For developers considering the model, that creates a practical division. Testing and fine-tuning can begin through Tinker, but any plan that requires direct access to the weights must wait for their publication.
Even then, the MoE architecture and the full model size will need to be considered together. The word "Small" should probably stay out of the hardware budget.