Vibeleaderboard

confirmed

Models

Updated Sep 22, 2026, 12:41 PM UTC

Xiaomi's MiMo V2.6 Lands on Vercel's AI Gateway

Two new MoE model tiers bring a 1M-token context and up to 128K output tokens to Vercel's AI Gateway for coding and agent workloads.

How this coverage works

This article combines reporting from 1 supporting Intel source. It is updated as material evidence arrives; prior published revisions remain in the record.

Xiaomi's MiMo V2.6 family, comprising Pro, Flash, and a faster Pro UltraSpeed variant, is now available through Vercel's AI Gateway, according to Vercel's official changelog. The models combine coding, reasoning, and tool use with native text, image, audio, and video understanding.

MiMo V2.6 Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters, of which 42 billion are activated per token, positioned for complex software-engineering and long-running agent work, Vercel's listing states. MiMo V2.6 Flash uses a smaller configuration, 309 billion total parameters with 15 billion activated per token, aimed at more routine multimodal automation.

The family supports a 1-million-token input context alongside up to 128,000 output tokens per response, according to Vercel. The Pro UltraSpeed variant serves the same Pro capabilities at up to 20 times the output speed, intended for latency-sensitive, interactive use.

Developers can select the models directly inside a coding agent by installing the latest Vercel CLI and running vercel ai-gateway setup, per Vercel's documentation. AI Gateway provides a single API across models with usage and cost tracking, retries, and failover, according to the company.

The three-tier lineup, a large Pro model for heavy agentic and coding work, a lighter Flash model for everyday automation, and an UltraSpeed variant for latency-sensitive interaction, gives developers a single-provider way to trade capability for speed and cost, according to Vercel's listing.

Update history1 updates

New facts extend one article instead of spawning duplicate write-ups across its source Intel pages.

  1. Sep 22, 2026, 12:41 PM UTC

    Revision 1 · initial

    Initial publication.