Xiaomi's MiMo V2.6 models with 1M-token multimodal context arrive on Vercel AI Gateway

What happened

Vercel has added Xiaomi's MiMo V2.6 family to its AI Gateway, making three variants available through a single unified API: MiMo V2.6 Pro, MiMo V2.6 Flash, and MiMo V2.6 Pro UltraSpeed. According to the Vercel blog, the models are aimed squarely at coding agents and workloads where agents run for extended periods.

The model family

Vercel describes MiMo V2.6 as a model that blends coding, reasoning, and tool-calling abilities with native handling of text, images, audio, and video. The context window spans 1M tokens, which Vercel positions as enough to fit long code repositories, accumulated tool traces, and work that spans multiple agent sessions. The model also supports structured outputs and can produce up to 128K output tokens in a single call.

The three tiers map to different needs:

• xiaomi/mimo-v2.6-pro is the larger sparse mixture-of-experts checkpoint, with 1.02T total parameters and 42B activated per token. Vercel recommends it for demanding software engineering tasks and long-running agent work.
• xiaomi/mimo-v2.6-flash is a smaller sparse mixture-of-experts model at 309B total parameters, activating 15B per token. Vercel frames it as the more efficient choice for day-to-day agent tasks and multimodal automation where peak capability matters less than speed and cost.
• xiaomi/mimo-v2.6-pro-ultraspeed serves the Pro checkpoint at up to 20 times its output speed, with the same capabilities, and is built for interactive use where latency matters.

Getting started

To use MiMo V2.6 in a coding agent, Vercel says developers should install the latest Vercel CLI (npm i -g vercel@latest), run vercel ai-gateway setup, and then select one of the three model IDs in their agent. The models can also be tried out directly in Vercel's model playground.

What the gateway adds

Beyond access to the new models, Vercel presents the AI Gateway as production infrastructure: a unified API for calling models from different providers, usage and cost tracking, and built-in handling of retries, failover, and performance optimizations, which the company says results in higher uptime than calling a model provider directly. The gateway also includes custom reporting, budgets that can be attached to individual API keys, and routing rules.

Why it matters

The sparse mixture-of-experts design is central to this release. Because Pro activates only 42B of its 1.02T parameters per token, it can carry the capacity of a very large model while running closer to the cost of a mid-sized one. Flash makes the same trade-off more aggressively at 15B activated parameters. For teams building agents that need to read entire repositories or carry long tool histories, a 1M-token multimodal context combined with a 128K output budget reduces the need to chunk, prune, or summarise state between steps.

The UltraSpeed variant targets a separate pain point. Agent loops that stream reasoning and generated code feel sluggish when output tokens trickle out, and an option to run a large checkpoint at up to 20 times the output speed makes it viable for interactive tools rather than just batch processing.

Finally, the arrival of Xiaomi-built models on Vercel's platform is a signal of how crowded the frontier model market has become. A company best known for phones and consumer electronics is now one API route away for developers already using the gateway, and for Vercel, each notable new model added strengthens the case for routing through a single provider-agnostic layer instead of managing separate relationships with many model providers.