Inkling 2 Small arrives on Vercel's AI Gateway for code agents with computer vision

Vercel announced in its changelog that Inkling 2 Small, a multimodal reasoning model from Thinking Machines, is now available on the AI Gateway. The model accepts text and image as input and returns text as output, which places it in a specific category within the Gateway's catalog: it's not just another text LLM, it's a model built for tasks that involve seeing something before acting.

Thinking Machines has been positioning its models as tools for professional knowledge work, not just chat. The name "Small" suggests there's a strategy of different sizes behind the Inkling 2 family, but Vercel doesn't detail larger variants in the changelog, so that remains an open point.

What it's for in practice

According to Vercel, Inkling 2 Small can work on longer coding tasks, use tools to search for information, and turn what it finds into documents, spreadsheets, and other deliverables. The visual part is what draws the most attention from people who work with front-end: the model reads text inside images, interprets charts and infographics, and inspects interfaces.

The most interesting use for UI developers is the visual development loop described by Vercel itself: the model can use screenshots to check the output of the code it wrote and guide the next revision. In practice, this means an agent capable of running a cycle of "wrote the component, took a screenshot, compared it to what was expected, adjusted," without a human needing to manually copy and paste a screenshot at every iteration.

How to call the model via AI SDK

The integration follows the pattern Vercel already uses for any Gateway model: swap the string, without changing the rest of the code. The model's name is thinkingmachines/inkling-2-small, and the example published in the changelog uses the streamText() function from the ai package:

import { streamText } from 'ai';

const result = streamText({
  model: 'thinkingmachines/inkling-2-small',
  prompt: 'Write a TypeScript function that groups tasks by status.',
});

for await (const text of result.textStream) {
  process.stdout.write(text);
}

Anyone who already uses Vercel's AI SDK for other models (GPT, Claude, Gemini) doesn't need to learn anything new: just swap the model value and, if using image as input, build the prompt with the multimodal content the API already accepts. This is the central point of the AI Gateway as a product: it abstracts the provider layer so that testing a new model costs one line of code, not a new integration.

Using it inside Claude Code, Codex, and Cursor

For those who prefer working inside a code agent already installed instead of calling the API directly, Vercel documents a three-step path:

1. Install the latest version of the Vercel CLI.
2. Run npx vercel ai-gateway setup in the project.
3. Select thinkingmachines/inkling-2-small as the model inside the agent (Claude Code, Codex, Cursor, among others).

This matters because it decouples the choice of model from the choice of tool: the same Cursor or Claude Code that today uses a model from Anthropic or OpenAI can, with this configuration, run specific tasks on Inkling 2 Small without switching editors or workflows. For those who test models side by side on visual coding tasks, this significantly reduces the friction of comparison.

What the AI Gateway solves behind the scenes

The Gateway isn't just a call router: it also handles billing, failover, and observability, which matters when you're running a new model in production without yet knowing how it behaves under load. The features cited by Vercel include:

• Usage and cost reports per API key.
• Support for Zero Data Retention, for those with compliance restrictions on retaining prompt data.
• Configurable budgets per key.
• Routing rules between providers, including automatic failover.

One point worth highlighting: Vercel states that the Gateway reflects the provider's price with no markup and charges no platform fee on inference, including on Bring Your Own Key (BYOK) requests. This changes the calculation for those who hesitated to use a gateway out of fear of paying twice for the same token, once to Thinking Machines and once to Vercel.

Where this helps and where it's still too early to trust

In short: Inkling 2 Small is one more piece on a board that already has dozens of models in the same Gateway, and the decision to use it should come from a real test on the specific task, not from novelty.

For automated UI review (taking a screenshot, comparing it to the design, suggesting CSS or markup adjustments), the model fills a niche that today is covered by visual regression testing tools or by generic multimodal models forced to fit that role. Having a model that already treats this use case as first-class is different from using a general-purpose model and hoping it understands the visual context.

What's still missing, because Vercel's changelog doesn't provide this data, is a quality benchmark compared to other vision models available in the same Gateway (such as those in the Claude or Gemini families) on interface inspection tasks. Before replacing an already-validated visual QA pipeline, the safer path is to run the same set of screens through the candidate models and compare accuracy, not to assume that "multimodal and built for code" already guarantees a better result in your specific case.

It's also worth noting that the model only produces text output: for workflows that depend on generating images (mockups, visual variations), it's not the right piece, even though it's good at interpreting input images. Anyone who needs visual generation still depends on another model in the same Gateway, possibly orchestrated alongside Inkling 2 Small in a chain of agents.