Meta's Muse Wonder model with 1M-token context arrives on Vercel's AI Gateway

Meta's Muse Wonder lands on Vercel's AI Gateway

Vercel has added Muse Wonder, a model from Meta built for agentic work, to AI Gateway, its unified API for calling large language models. According to a post on the Vercel blog, the model is available now and can be used by anyone holding an AI Gateway key, with no separate Meta account required.

A model aimed at long-running agents

Vercel positions Muse Wonder around three related workloads: coding, long-running agents and computer use. These are scenarios where a model has to sustain a task across many steps rather than answer a single prompt. The Vercel blog says the model can break a job into steps, hand parts of it off to sub-agents, and recover from errors as a run continues instead of abandoning the task the moment something breaks.

The headline specification is a 1M-token context window. The model also accepts text, image and PDF input, supports tool calling, and can return structured output. Because these capabilities share a single context, Vercel notes that an agent can move between code, screenshots and documents within one workflow — the combination a computer-use agent needs when it has to reason about what is on screen while also reading or editing files.

How to start using it

Setup runs through the gateway rather than through Meta directly. Developers set AI_GATEWAY_API_KEY in their environment and reference the model as meta/muse-wonder-1.0. For coding agents, Vercel offers a guided path: run npx vercel ai-gateway setup and select meta/muse-wonder-1.0 in the agent's model settings. The model can also be tried in the AI Gateway model playground or browsed in the model catalog.

Why it matters

Two things make this release notable. The first is distribution: reaching a Meta model through Vercel's single API removes the account setup and key management that comes with going to each provider separately. As Vercel presents it, AI Gateway is one interface for invoking models, tracking usage and spend, and configuring routing, retries and failover — with the option to use an AI Gateway key or bring your own provider key.

The second is the workload the model targets. Long-running agents tend to fail for mundane operational reasons: a step errors out, a transcript outgrows the context, or a provider hiccups mid-run. A 1M-token window, built-in error recovery and the ability to split work across sub-agents address the first kind of problem, while gateway-level retries and failover cover the second. For teams already running agent loops on Vercel, Muse Wonder is now effectively a one-line model swap away.