MiMo V2.5 Pro is Xiaomi’s most capable AI model, released April 22, 2026. It is built specifically for long-horizon agentic tasks: the kind of work where a model needs to make hundreds or thousands of tool calls in a single session, manage its own context, and keep going for hours without losing track of what it’s doing. The headline numbers are 57.2% on SWE-bench Pro, 64% Pass^3 on ClawEval, and 40-60% fewer tokens than Claude Opus 4.6 at comparable capability levels.

That last point is the real story. MiMo V2.5 Pro doesn’t just match frontier models on agentic coding benchmarks. It does so while burning dramatically fewer tokens per task. If you’ve been watching your API bills climb with Opus 4.6 or GPT-5.4 on complex agent workflows, this model changes the math.

Here’s everything you need to know.

Architecture

MiMo V2.5 Pro shares the same base architecture as MiMo V2 Pro but ships with significant training improvements that push it into a different performance tier. The foundation is a Mixture-of-Experts transformer with over 1 trillion total parameters, where 42 billion activate per token.

The 1-million-token context window is paired with a hybrid attention mechanism that mixes dense and sparse attention patterns. This lets the model handle extremely long sessions without the quality degradation you typically see when models approach their context limits. For agentic tasks that run for hours and accumulate thousands of tool call results, this matters more than raw context length.

The MoE architecture keeps inference costs manageable. Only 42B parameters fire per token despite the 1T+ total, which means you get frontier-level reasoning without frontier-level compute requirements on the provider side.

Key Capabilities

Long-Horizon Agentic Tasks

The defining feature of V2.5 Pro is its ability to sustain 1,000+ tool calls in a single session. Most models start degrading well before that point. They lose track of earlier context, repeat work, or make contradictory decisions. V2.5 Pro was trained specifically to avoid these failure modes.

Two demonstrations from Xiaomi’s release showcase this:

SysY Compiler in Rust. V2.5 Pro built a complete SysY compiler in Rust in 4.3 hours. It made 672 tool calls across the session and passed 233 out of 233 tests. That’s not a toy project. A SysY compiler involves lexing, parsing, semantic analysis, and code generation, all in a systems language with strict ownership rules.

Video Editor (8,192 lines). V2.5 Pro built an 8,192-line video editor in 11.5 hours using 1,868 tool calls. This is the kind of project that would take a human developer days or weeks. The model planned the architecture, implemented features incrementally, and tested as it went.

These aren’t cherry-picked demos of simple code generation. They’re sustained, multi-hour engineering sessions where the model acts as an autonomous developer.

Harness Awareness

V2.5 Pro introduces what Xiaomi calls “harness awareness.” The model understands the tool-calling environment it’s operating in and actively manages its own context and memory. Instead of blindly consuming context until it runs out, V2.5 Pro makes strategic decisions about what to keep, what to summarize, and when to offload information.

This is a meaningful shift from how most models handle agentic workflows. Typically, the harness (Claude Code, Aider, OpenCode, etc.) manages context on behalf of the model. V2.5 Pro participates in that management, which leads to better decisions about what information matters at each step.

Token Efficiency

This is the headline feature. On ClawEval, V2.5 Pro uses 40-60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 while achieving comparable or better results. It scores 64% Pass^3 on ClawEval with approximately 70K tokens per trajectory.

To put that in perspective: if Opus 4.6 needs 150K-175K tokens to complete a complex agentic task, V2.5 Pro does it in roughly 70K. That’s not a minor optimization. It cuts your API costs in half or better, and it means faster task completion since fewer tokens means less time waiting for generation.

Xiaomi positions V2.5 Pro in the “upper-left corner” of the capability-vs-cost chart. Same capability, dramatically fewer tokens. For teams running agentic workflows at scale, this is the most important metric.

Benchmarks

The SWE-bench Pro score of 57.2% puts V2.5 Pro at the top of the leaderboard for this benchmark. But the ClawEval results tell the more interesting story. V2.5 Pro doesn’t just score well. It scores well while using a fraction of the tokens that competing models need.

Kimi K2.6 is the closest competitor on token efficiency, but V2.5 Pro still uses roughly 40% fewer tokens while scoring higher on Pass^3.

Token Efficiency: The Real Story

If you’re evaluating V2.5 Pro against other frontier models, token efficiency is where you should focus. The raw benchmark scores are competitive but not dramatically ahead. The token usage is dramatically ahead.

Here’s why this matters in practice:

Cost. If you’re running agentic workflows through an API, you pay per token. A model that uses 40-60% fewer tokens to achieve the same result cuts your bill by 40-60%. For teams running hundreds of agent sessions per day, this adds up fast.

Speed. Fewer tokens means faster completion. Token generation is the bottleneck in most agentic workflows. If V2.5 Pro generates 70K tokens where Opus 4.6 generates 160K, your task finishes in roughly half the time.

Context headroom. Agentic tasks accumulate context over time. A model that’s more efficient with tokens has more room to work before hitting context limits. This is especially important for the 1,000+ tool call sessions that V2.5 Pro is designed for.

Reliability. Models that generate fewer tokens tend to make fewer errors. There’s less opportunity for the model to contradict itself or go off track when it’s being concise and deliberate.

The combination of these factors makes V2.5 Pro particularly attractive for production agentic workflows where you’re optimizing for cost, speed, and reliability simultaneously.

Pricing

V2.5 Pro ships at the same base price as MiMo V2 Pro. No price increase despite the significant capability jump. But the effective cost is roughly half of what V2 Pro was, thanks to changes in the Token Plan:

• Context-length multiplier removed. V2 Pro charged more for longer contexts. V2.5 Pro doesn’t. You pay the same rate whether you’re using 10K or 500K tokens of context.
• Night-time discounts. Reduced rates during off-peak hours. If you’re running batch agent jobs, schedule them at night.
• Auto-renewal. Token Plans now auto-renew, so you don’t lose unused capacity or forget to top up mid-workflow.

The net effect is that V2.5 Pro costs roughly half what V2 Pro did for equivalent workloads, and it uses 40-60% fewer tokens on top of that. If you were spending $100/day on V2 Pro agent workflows, you might spend $25-30/day on V2.5 Pro for the same output.

For exact pricing, check platform.xiaomimimo.com. Pricing varies by region and plan tier.

How to Access MiMo V2.5 Pro

API Access

The primary way to use V2.5 Pro is through the API at platform.xiaomimimo.com. The model tag is mimo-v2.5-pro.

You can also access it through Xiaomi’s AI Studio interface for interactive use.

Using with Coding Harnesses

V2.5 Pro works with the major agentic coding tools:

• Claude Code. Configure as a custom provider pointing to the MiMo API endpoint. V2.5 Pro’s harness awareness means it works particularly well with Claude Code’s tool-calling patterns.
• OpenCode. Add the MiMo API as a provider in your OpenCode config. The model handles OpenCode’s file editing and terminal tool calls natively.
• Kilo. Full support as a backend model. Kilo’s lightweight harness pairs well with V2.5 Pro’s token efficiency.

If you’ve previously set up MiMo V2 Pro with Aider, the same configuration approach works for V2.5 Pro. Just update the model tag to mimo-v2.5-pro.

Quick Setup

# Set your API key
export MIMO_API_KEY="your-key-here"

# Example: using with OpenCode
opencode --provider mimo --model mimo-v2.5-pro

# Example: API call
curl https://api.xiaomimimo.com/v1/chat/completions \
  -H "Authorization: Bearer $MIMO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mimo-v2.5-pro",
    "messages": [{"role": "user", "content": "Your prompt here"}]
  }'

V2.5 Standard vs V2.5 Pro

Xiaomi released two models in the V2.5 family. They serve different purposes.

V2.5 Standard is the multimodal model. It handles image, audio, and video inputs alongside text. It’s faster and cheaper than Pro, making it the right choice for general-purpose tasks, content generation, analysis, and anything involving non-text modalities.

V2.5 Pro is the specialist. It’s built for complex, multi-step coding and agentic tasks where you need the model to sustain focus over hundreds or thousands of tool calls. If you’re building autonomous coding agents or running long-horizon engineering workflows, Pro is the one you want.

For most developers, the decision is straightforward: use Standard for everyday tasks and Pro for serious agentic coding work. They share the same API platform, so switching between them is just a model tag change.

Open Source

Xiaomi has confirmed that V2.5 Pro will be released as open source. No exact date yet, but the V2 family has a track record of open releases. MiMo V2 Flash was open-sourced and can be run locally, so there’s precedent.

When the open-source release happens, expect it to be available through the usual channels: Hugging Face, ModelScope, and Ollama. Given the 1T+ parameter count, you’ll need significant hardware to run the full model locally, but quantized versions should be more accessible.

For now, the API is the only way to use V2.5 Pro.

How V2.5 Pro Fits the Landscape

The AI model landscape in 2026 is crowded at the top. Claude Opus 4.6, GPT-5.4, Gemini 3.1 Pro, and Kimi K2.6 all compete for the “best agentic coding model” title. V2.5 Pro doesn’t try to win by being dramatically smarter than all of them. It wins by being dramatically more efficient.

This positions it well for teams that are past the “which model is smartest” phase and into the “which model gives us the best results per dollar” phase. If you’re running AI coding tools in production, cost and speed matter as much as raw capability.

Among Chinese AI models, V2.5 Pro represents a significant step forward for Xiaomi. The V2 family already established MiMo as a serious contender. V2.5 Pro puts it at the frontier.

FAQ

Is MiMo V2.5 Pro better than Claude Opus 4.6?

On SWE-bench Pro, yes: 57.2% vs roughly 55%. On ClawEval, the scores are close, but V2.5 Pro uses 40-60% fewer tokens to get there. Whether “better” means higher scores or better cost-efficiency depends on your priorities. For most production use cases, the token efficiency advantage makes V2.5 Pro the more practical choice for agentic workflows.

Can I run MiMo V2.5 Pro locally?

Not yet. The model is currently API-only through platform.xiaomimimo.com. Xiaomi has confirmed an open-source release is coming, but no date has been announced. When it drops, the 1T+ parameter count means you’ll need serious hardware for the full model. Quantized versions will likely be more practical for local use.

What’s the difference between V2.5 Pro and V2 Pro?

Same base architecture, but V2.5 Pro has significant training improvements. The practical differences: better long-horizon task performance (1,000+ tool calls vs hundreds), harness awareness for context management, 40-60% better token efficiency, and higher benchmark scores across the board. Pricing is the same or lower thanks to Token Plan changes.

Should I use V2.5 Standard or V2.5 Pro?

Use Standard for general-purpose tasks, especially anything involving images, audio, or video. Use Pro for complex coding tasks, autonomous agent workflows, and anything requiring sustained multi-step reasoning over hundreds of tool calls. Standard is faster and cheaper. Pro is more capable for its specific niche.