Xiaomi MiMo‑V2.5 Series Launches Public Beta with Stronger Agent and Multimodal Capabilities

Today Xiaomi announced the public beta of its MiMo‑V2.5 series, which comprises MiMo‑V2.5, MiMo‑V2.5‑Pro, MiMo‑V2.5‑TTS, and MiMo‑V2.5‑ASR. The release emphasizes stronger reasoning, more stable agents, longer context windows, better instruction following, and improved full‑modal perception, marking a shift from merely usable to truly useful.

MiMo‑V2.5‑Pro: Advanced Agent Capabilities

MiMo‑V2.5‑Pro is described as Xiaomi's most powerful model to date. In general‑purpose agent ability, complex software engineering, and long‑range tasks it can contend with leading agents such as Claude Opus 4.6 and GPT‑5.4, achieving a comprehensive upgrade over the previous MiMo‑V2‑Pro.

Internal testing showed that, when paired with an appropriate runtime framework, the model reliably completed a single task involving nearly a thousand tool calls, demonstrating markedly improved instruction‑following, implicit requirement capture, and logical consistency over extended periods, allowing it to undertake serious professional work with higher confidence.

Example: Building a Complete SysY Compiler in Rust

The task originated from Peking University’s "Compiler Principles" course, requiring a full SysY compiler written in Rust—including lexer, parser, AST, Koopa IR generation, RISC‑V backend, and performance optimizations. Undergraduate students typically need several weeks; MiMo‑V2.5‑Pro completed it in 4.3 hours, executing 672 tool calls and achieving a perfect 233/233 score on the hidden test set, showcasing high productivity.

The model did not rely on brute‑force trial‑and‑error. It first constructed the entire pipeline skeleton, then tackled each layer step‑by‑step: Koopa IR (110/110), RISC‑V backend (103/103), and performance optimization (20/20). The first compilation passed 137/233 tests, a 59 % cold‑start pass rate, indicating a correct architecture before any test execution. At iteration 512, a refactor caused two test regressions, which the model diagnosed, recovered, and continued progressing.

Example: Developing a Web‑Based Video Editor

With a simple prompt—"build a video editor web application"—MiMo‑V2.5‑Pro delivered a runnable web app featuring multi‑track timelines, clip trimming, cross‑fade, audio mixing, and export workflows. The final codebase comprised 8,192 lines, required 1,868 tool calls, and was completed autonomously in 11.5 hours.

Agent and Multimodal Advances Across the Series

MiMo‑V2.5 is a native multimodal large model designed for agent scenarios, capable of seeing, hearing, and reading, then converting understanding into actions. In authoritative agent benchmarks such as Claw‑Eval, MiMo‑V2.5 surpasses MiMo‑V2‑Pro, handling everyday tasks while reducing API cost by roughly 50 %.

Multimodal perception also outperforms the previous MiMo‑V2‑Omni, with gains in cross‑modal reasoning, video understanding, and chart analysis. Evaluations on VideoMME, CharXiv, and MMMU‑Pro show performance approaching or exceeding leading closed‑source models.

Token Efficiency and Pricing Optimizations

MiMo‑V2.5‑Pro saves 42 % of tokens compared with Kimi K2.6.

MiMo‑V2.5 saves 50 % of tokens compared with Muse Spark.

The new Token Plan introduces a 1× credit rate for MiMo‑V2.5 (1 Token = 1 Credit) and a 2× rate for MiMo‑V2.5‑Pro (1 Token = 2 Credits). The previous 1 Token = 4 Credits for cancelled tokens is removed, and the credit multiplier no longer distinguishes between 256 k and 1 M context windows.

Night‑time discounts (00:00–08:00 BJT) apply an additional 20 % off the existing rate. New subscription options include auto‑renewal monthly plans (7 % off the next month) and annual plans (12 % off the whole year), both limited to a single use.

Upcoming Open‑Source Release

MiMo‑V2.5‑Pro and MiMo‑V2.5 are slated for global open‑source release in the near future.

官方API接入
Xiaomi MiMo 开放平台：https://platform.xiaomimimo.com
Token Plan 订阅：https://platform.xiaomimimo.com/#/token-plan

产品体验
Xiaomi MiMo Studio：https://aistudio.xiaomimimo.com/#/c