Vercel adds Thinking Machines' Inkling 2 Small to its AI Gateway

Inkling 2 Small lands on Vercel's gateway

Vercel has added Inkling 2 Small, a multimodal reasoning model from Thinking Machines, to the set of models reachable through its AI Gateway. In a changelog entry published on 13 November, Vercel describes the model as one that takes both text and images as input, produces text output, and is aimed at developers building coding agents as well as at visual understanding and broader knowledge work.

What the model can do

The model is positioned for long-horizon programming work. According to Vercel, it can sustain extended coding tasks, call tools to pull in information it needs, and shape the results into concrete outputs such as documents or spreadsheets.

Its image handling covers several practical cases: reading text embedded in pictures, interpreting charts and infographics, and reviewing user interfaces. The last capability feeds a notable development workflow — the model can examine screenshots of the software it has just written, judge whether the output looks right, and use that judgement to guide the next revision. That closes the loop between generating code and verifying what it actually renders, without a human having to describe the screen back to the model.

How to get started

The model is exposed under the identifier thinkingmachines/inkling-2-small. To wire it into agent tools such as Claude Code, Codex or Cursor, Vercel instructs users to install the latest Vercel CLI and run its setup command:

npx vercel ai-gateway setup

The model can then be selected inside the agent. Vercel also points to its coding agents guide for further configuration details, and offers a model playground for anyone who wants to evaluate Inkling 2 Small before committing to an integration.

What the gateway layer provides

The release also highlights what Vercel's gateway is for. According to the announcement, AI Gateway sits as a unified API in front of many models and adds usage and cost tracking, retry and failover handling, and performance tuning that Vercel says delivers better uptime than calling a provider directly. Further features include built-in custom reporting, Zero Data Retention support, per-API-key budgets and routing rules.

On pricing, Vercel states that the gateway mirrors provider rates with no markup and collects no platform fee on inference — including for Bring Your Own Key requests, where customers route calls through their own provider credentials.

Why it matters

Two things make this release worth noting. First, it gives developers another option in the coding-agent niche through an API many teams already use, alongside the other models listed on the gateway. The screenshot-driven revision loop in particular targets a real gap in agentic development: checking rendered output rather than trusting that code works on paper.

Second, the gateway's pricing posture — provider rates, no inference fee, BYOK-friendly — lowers the cost of experimenting with a newer model, since switching amounts to changing a model string in a config. As agent frameworks increasingly standardise around provider-agnostic model selection, distribution channels like this one will help decide which new models actually get tried in production.