MiniMax M3 and M2.7 run free on Vercel AI Gateway through September 6

Free access through September 6

Vercel has opened free access to MiniMax's M3 and M2.7 models through its AI Gateway. According to the Vercel blog, the offer is served by GMI Cloud and runs through Sunday, September 6. During that window, requests to two dedicated model identifiers — minimax/minimax-m3-free and minimax/minimax-m2.7-free — are routed to GMI Cloud at no charge.

There is one important caveat: Vercel says the -free identifiers will return errors once the promotion ends. They will not silently convert into paid usage, so anything still pointing at them after September 6 will simply fail.

Two ways to route requests

Developers can take one of two approaches. The first is to call the -free model IDs directly, which is the quickest path for experimentation. Vercel demonstrates this with an AI SDK example that passes minimax/minimax-m3-free as the model in a streamText call with a straightforward prompt.

The second approach is aimed at code that needs to keep working past the promotion. Instead of using the -free ID, you keep the standard model identifier, such as minimax/minimax-m3, and configure GMI Cloud as the preferred provider through the gateway's providerOptions, setting order: ['gmicloud']. In this setup, the gateway falls back to another provider whenever GMI Cloud cannot serve a request, so the application keeps running after the free window closes — traffic just moves to another serving path.

According to Vercel, both options work across every AI Gateway API, so the choice is not limited to projects built on the AI SDK.

Wiring the models into coding agents

The announcement also covers coding agents. Vercel points developers to its coding agents guide and provides a setup command, vercel ai-gateway coding-agents setup, which connects tools including Claude Code, Codex, OpenCode, Cursor and Pi to the gateway. Once an agent is connected, users can select minimax/minimax-m3-free from inside the agent and run coding tasks against the free model.

Why it matters

Evaluating an unfamiliar model normally involves signing up with another provider, adding a payment method and writing integration code before a single test can run. A time-boxed free window on a gateway that many developers already have wired into their applications removes most of that friction: swap in a model string and the experiment begins.

For MiniMax, the promotion is a distribution play — it puts M3 and M2.7 in front of developers who might otherwise never try them. For developers, it is a low-risk chance to benchmark the models against the ones they already use, including in coding-agent workflows, where differences in model quality tend to surface quickly.

The main operational detail is expiry. Because the -free identifiers hard-fail after September 6, teams that find the models useful should plan a switch to the standard IDs, ideally with a provider order configured so traffic fails over gracefully rather than breaking outright.