Xiaomi’s MiMo AI division, led by former DeepSeek researcher Luo Fuli, released four new models on April 23, 2026 — the MiMo-V2.5-Pro flagship, the MiMo-V2.5 multimodal agent model, and a complete voice suite comprising MiMo-V2.5-TTS Series and MiMo-V2.5-ASR. The release came just 36 days after the previous MiMo-V2 series launch. On the Artificial Analysis Intelligence Index, MiMo-V2.5-Pro ranked first among all open-source large models in overall intelligence, and also tied for first in the Agent-specific index. It additionally placed in the global top five across both open-source and closed-source models. This article covers the specs, benchmarks, real-world task results, and what each new model in the V2.5 series actually does — including how it compares to Meta’s AI models and other frontier labs.
From a Phone Maker to Open-Source AI’s Front Row
Xiaomi is better known for smartphones and smart home hardware, but since hiring Luo Fuli from a competitive Chinese AI talent market in late 2025, its MiMo division has been building at a speed that has drawn attention from developers worldwide. Luo previously contributed to DeepSeek’s V2 model and is colloquially described in Chinese AI circles as “the genius girl.” When MiMo-V2-Pro was released as an anonymous model called Hunter Alpha on OpenRouter on March 11, 2026, it topped usage charts for several days and processed over one trillion tokens before Xiaomi confirmed the model’s identity on March 18.
MiMo-V2.5-Pro entered public beta on April 22–23, 2026, followed the same day by the release of the V2.5 multimodal agent model. On April 24, Xiaomi released MiMo-V2.5-TTS Series and MiMo-V2.5-ASR, completing a full voice input-output pipeline. The timing is notable: DeepSeek V4 had been rumoured for release in the same week, and MiMo-V2-Pro had already been mistaken for it once.
Live Benchmarks
How MiMo-V2.5-Pro Scores
Select a benchmark to compare scores. All figures are from official Xiaomi data or verified third-party sources.
Source: Xiaomi MiMo official release, April 23 2026. MiMo Coding Bench is Xiaomi’s internal evaluation suite covering repo understanding, project building, code review, and software engineering tasks.
SWE-bench Pro tests models on real software engineering tasks pulled from production open-source repos. Figures are verified by Artificial Analysis as of April 23, 2026.
ClawEval is an end-to-end agent evaluation benchmark measuring multi-step task completion. MiMo-V2.5-Pro also consumes 42% fewer tokens than Kimi K2.6 at equivalent ClawEval score. Source: Xiaomi MiMo platform.
Humanity’s Last Exam tests expert-level reasoning across domains. Figures sourced from Xiaomi’s official April 23, 2026 release data.
Token efficiency is one of the bigger talking points in the V2.5 release. At the same ClawEval benchmark score, MiMo-V2.5-Pro uses 42% fewer tokens than Kimi K2.6 (Moonshot AI’s open-source multimodal agent model released earlier in April 2026), and MiMo-V2.5 uses 50% fewer tokens than Meta Maverick. For developers running agentic workflows at scale, token volume directly translates to operating costs — so these comparisons carry practical weight beyond benchmark positioning.
MiMo-V2.5-Pro is built on a Mixture-of-Experts (MoE) architecture with over one trillion total parameters and 42 billion active per inference pass. It supports up to a 1M-token context window. Xiaomi’s internal tests show it can stably complete single tasks involving nearly a thousand rounds of tool calls, which covers scenarios like running the entirety of a complex software project from scaffolding to delivery. On the competitive 2026 AI model landscape, this kind of long-horizon agentic capability is where the major labs are competing most directly.
MiMo Milestones
From Zero to #1 Open Source
The MiMo series moved from a 7B reasoning model to a top-ranked 1T-parameter agent model in under a year. Tap each milestone for more detail.
What’s New
The V2.5 Model Lineup
Four models released across two days. Each covers a different part of the AI agent pipeline.
Xiaomi’s most capable model. Designed for long-horizon and difficult agent tasks. Can stably complete single tasks with nearly a thousand rounds of tool calls. In public beta as of April 22, 2026; open source release planned.
Native full-modal agent model covering images, audio, and video simultaneously. Faster average inference speed than the Pro version — suited for latency-sensitive tasks. Agent capabilities surpass the previous flagship MiMo-V2-Pro, while cutting API costs by ~50% vs Meta Maverick.
Three distinct TTS models sharing unified style-instruction following and audio-label control. Available free for a limited time on the MiMo Open Platform.
Fully open-sourced on April 24, 2026. Supports Chinese dialects (Wu, Cantonese, Minnan, Sichuan), English, Chinese-English code-switching, strong noise, and multi-speaker scenarios. Outputs punctuation natively so transcripts can be used without post-processing.
Interactive Tool
Token Cost Calculator
Estimate the difference in token usage between MiMo-V2.5-Pro and comparable models at scale. Figures are based on official Xiaomi comparative data from the ClawEval benchmark at equivalent performance.
Based on Xiaomi’s official comparative data from the ClawEval agentic benchmark. Token savings percentages reflect same-score comparisons, not identical workloads.
Real-World Tasks
What MiMo-V2.5-Pro Actually Built
Xiaomi published three long-horizon task results. These are not simple prompts — they are multi-hour, multi-step autonomous engineering runs.
The model built the full compiler pipeline layer by layer — lexer, parser, AST, Koopa IR code generation, RISC-V assembly backend, and performance optimisation. The first compile passed 59% of tests cold, meaning the architecture was correct before any iterative fixing. Full marks on Koopa IR, RISC-V backend, and performance optimisation. PKU CS students typically take several weeks to complete this course project. MiMo-V2.5-Pro finished it in 4.3 hours with a perfect hidden test score of 233/233.
The model ran for 11.5 hours autonomously, making 1,868 tool calls, and produced 8,192 lines of code. The delivered application includes multi-track timeline, segment cropping, cross-fading transitions, audio mixing, and export flow — a feature set that would constitute a substantive web-based video editing product.
The task required determining power transistor size, adjusting a compensation network, and selecting a bias voltage so that six circuit specifications — phase margin, line adjustment rate, load adjustment rate, static current, power suppression ratio, and transient response — were all met simultaneously. Experienced analog circuit designers typically spend days on such a project. Connected to an ngspice simulation loop using Claude Code as the simulation framework, MiMo-V2.5-Pro completed a design meeting all six target metrics in approximately one hour, with four key metrics an order of magnitude better than its initial design.
MiMo-V2.5-Pro benchmark comparison — Artificial Analysis Intelligence Index, April 2026 | Photo Source: Xiaomi MiMo Platform / Artificial Analysis
Voice Pipeline
The Speech Layer: TTS Series + ASR
Released April 24, 2026. Three text-to-speech models and one open-source speech recognition model complete the full voice input-output chain.
Built-in high-quality voices available out of the box. Supports fine-grained control over speech rate, emotion, and tone via natural language instructions. Operates without additional setup.
Generate a new voice from a single natural language description — no reference audio needed. Supports age, gender, accent, timbre, and personality descriptors as generation parameters.
Reproduces a target voice from as little as a few seconds of reference audio with no fine-tuning or annotation required. Cloned voices retain all style-instruction and audio-tag control capabilities of the series.
ASR model weights and code are open-sourced at github.com/XiaomiMiMo/MiMo-V2.5-ASR. TTS Series is available for free trial at Xiaomi MiMo Open Platform.
Alongside the model releases, Xiaomi also updated its MiMo Token Plan subscription structure. The old system billed one token as four credits; that distinction has been removed. The new system bills MiMo-V2.5 at 1×(one token = one credit) and MiMo-V2.5-Pro at 2× (one token = two credits). An exclusive night rate applies daily from 00:00 to 08:00, offering an additional 20% discount on all model credits. Annual subscribers receive a further reduction: the annual subscription saves 948.96 yuan compared to monthly billing. All eight current MiMo series models are covered under the Token Plan, including the newly launched TTS and VoiceClone models.
The release sits within a broader AI development push by Xiaomi. In March 2026, CEO Lei Jun announced plans to invest at least US$8.7 billion in artificial intelligence over the following three years. The MiMo division’s output — going from a 7B reasoning model in April 2025 to an open-source #1 ranking in April 2026 — has been one of the more closely watched trajectories in Chinese AI. Related context on AI integration with hardware and devices is covered in our reporting on AI-powered device capabilities in 2026 and the leadership shifts at major tech companies reshaping the industry.
Token Plan Update
MiMo Token Plan: What Changed
The updated subscription plan covers 8 MiMo models. Credits billing was simplified and night-time discount rates were added.
| Model | Old Rate | New Rate |
|---|---|---|
| MiMo-V2.5 | — | 1× (1 Token = 1 Credit) |
| MiMo-V2.5-Pro | 4× Credits | 2× (1 Token = 2 Credits) |
| Night rate (00:00–08:00 daily) | Standard rate | Additional 20% off all models |
| Annual subscription saving | — | 948.96 yuan vs monthly billing |
Context and 1M context windows are no longer billed at different Credit multipliers. Old users activating auto-renewal receive 70% off the following month; new users receive 77% off. Source: Xiaomi MiMo official release.
This article covered the MiMo-V2.5 series launch, including benchmark performance across SWE-bench Pro, ClawEval, MiMo Coding Bench, and Humanity’s Last Exam; three real-world long-horizon task results from the PKU compiler project, the video editor build, and the analog EDA circuit design; details on all four new models in the V2.5 family; the complete MiMo-V2.5-TTS and ASR voice pipeline; and the updated Token Plan pricing structure. The models are accessible via the Xiaomi MiMo Open Platform, with ASR weights and code available on GitHub and HuggingFace.
