Multimodal · NOVEMBER 8, 2026
Inkling 2 Small can visually refine websites—but launches with closed weights
Thinking Machines’ new multimodal model offers a hosted route to website prototyping. Its visual-feedback loop warrants testing, not assumptions about better conversions.
Thinking Machines launched Inkling 2 Small with hosted access but no downloadable weights. VentureBeat reports that Thinking Machines launched Inkling 2 Small on November 5, making a mixture-of-experts model with 594 billion total parameters and 17 billion active parameters available through Tinker.
The announcement and Artificial Analysis’s launch evaluation establish the boundary: access starts with hosted inference and fine-tuning, not downloadable weights. The website-prototyping proposition comes from Thinking Machines: the company says it can use visual feedback to build and refine websites. That’s a capability claim, not evidence of better conversions.
Visual feedback needs a supervised website test
The company describes the image pathway: image inputs pass through a 292-million-parameter vision encoder, with image patches interleaved with text tokens. Its announcement also specifies the intended feedback loop: "The model uses vision in the loop to interpret and validate the visual coding output it generates."
Thinking Machines reports a self-reported score of 1,340 on its DesignArena Website snapshot dated November 5. The secondary coverage adds context: The Decoder also highlights that website score. But the company’s qualification matters: its claim to lead open-weights models is explicitly conditional on releasing the weights. As the announcement and The Decoder make clear, the launch remains proprietary.
For a small-business owner, the testable workflow is narrow: prototype landing-page or product-page variants, review rendered pages and factual claims, then measure qualified inquiries or purchases. Company-selected demonstrations don’t establish customer-acquisition gains.
A supervised trial should precede adoption.
A supervised trial should precede adoption.
Benchmark scores, token use, and the weights boundary
The independent evaluations provide different measurements. Artificial Analysis gives the model an Intelligence Index score of 31. The verified ARC Prize result is separate: 81.7% on ARC-AGI-2 semi-private at XHigh reasoning. Neither result establishes how well a particular business website will perform.
Generation volume also deserves attention. According to the evaluator, Inkling 2 Small averages roughly 150,000 output tokens per Intelligence Index task, versus roughly 37,000 for Inkling Small. Those are benchmark workloads, not website-project budgets. Operators shouldn’t treat them as estimates for building a landing page.
The next access change remains prospective. The company says Apache 2.0 weights are intended for the week after launch, following external safety testing and red-teaming. It also specifies the checkpoint implication: customized checkpoints become downloadable with that release.
That distinction is reflected in the reporting: VentureBeat likewise distinguishes hosted access from future self-hosting. For now, the decision is whether to test the hosted workflow through Tinker, not whether to deploy downloadable weights. The planned release will change that boundary only when it happens.