Contents
Thinking Machines released Inkling 2 Small on November 5, 2026, making inference and fine-tuning available through Tinker. Full model weights are scheduled for release next week, allowing developers to try the API now and review deployment requirements before considering self-hosting.
Key Points#
- Inkling 2 Small supports text and image inputs and activates 17 billion parameters per token out of a total of 594 billion in its mixture-of-experts model.
- Its reported AA Intelligence Index score is 31, up from 13 for the previous Inkling Small.
- Inference and fine-tuning are available through Tinker, but the full model weights remain unreleased as of November 7.
- Published benchmarks show competitive coding performance, though it does not consistently outperform the comparison models.
What Is Available Now#
Developers can use Inkling 2 Small for inference and fine-tuning through Tinker, Thinking Machines’ API service. Teams can evaluate the base model, then test whether additional training improves performance on specific tasks.
The model uses a mixture-of-experts (MoE) architecture. It has 594 billion parameters in total, but only 17 billion are active per token. This distinction matters: the active parameter count indicates how much of the model participates in computing each token, while the total parameter count affects storage and deployment of the entire model.
It supports image inputs alongside text. For developers building tools that inspect screenshots or other visual material, it is worth evaluating on workflows that require both reading comprehension and visual inspection.
Potential Uses#
According to Thinking Machines, Inkling 2 Small is particularly strong at tasks involving many sequential steps. One example is building a website, checking its appearance, and making corrections. The company also says the model can adapt to different harnesses—the software that manages tools, context, and progress—and execution environments.
Another demonstrated use involved having Inkling 2 Small fine-tune another model through Tinker. This suggests a role in automating parts of model development, but the demonstration alone does not establish reliability on unfamiliar datasets or training objectives.
Practical evaluations should therefore use complete workflows. For a website task, measure whether the model can complete the requested changes, identify visual defects, and fix them without introducing new problems. A plausible first response provides less convincing evidence than a completed and verified deliverable.
Comparing Scores#
The announcement reports an AA Intelligence Index version 5.0.1 score of 31, up from 13 for the previous model. The comparison table, however, uses AA Index v5.0 and lists Inkling 2 Small at 30.6. These versions need to be distinguished when comparing results.
The following figures are taken from the announcement. Higher scores are better.
| Model | AA Index v5.0 | FrontierCode v2.0 | DeepSWE v1.1 | |---|---:|---:|---:| | Inkling 2 Small | 30.6 | 25.4% | 72.1% | | DeepSeek-V4.1-Flash-1017 | 32.6 | 28.5% | 77.0% | | GPT-6 Luna | 27.9 | 25.2% | 66.6% | | Claude Haiku 5.5 | 36.5 | 34.9% | 59.9% |
In the comparison presented, Inkling 2 Small trails DeepSeek-V4.1-Flash-1017 and outperforms GPT-6 Luna and GLM-5.3-Flash across all four metrics, including Vals Index v2.2. Results against Claude Haiku 5.5 are mixed: it scores lower on the two overall indices and FrontierCode, but higher on DeepSWE. These results justify trying it for coding tasks, but do not establish a ranking that applies to every use case.
What Still Needs Checking#
According to Thinking Machines’ Soumith Chintala, pretraining, mid-training, and post-training used NVIDIA’s NVFP4 four-bit floating-point format. He said the model underwent 68 hours of pretraining on NVIDIA GB300 infrastructure, followed by nine days of post-training involving 17 million rollouts. These are company-reported training figures, not measurements of customer inference costs or speed.
Chintala said the weights would be released next week after external safety testing is complete. As of November 7, the release is still planned. Even when the weights become available, their availability alone will not clarify licensing terms, supported deployment software, or hardware requirements; these will need to be checked at release.
Before adopting it in production, check Tinker’s pricing, latency, usage limits, and performance on your own inputs. The benchmark data presented does not answer these operational questions.
Next Steps#
Run one representative multistep task with Inkling 2 Small on Tinker, and compare the quality of the deliverable, total cost, and time required against your current model.
Source: https://gihyo.jp/article/2026/11/inkling-2-small
Contents
