MiMo-V2.5-Pro API

Technical Specifications of MiMo-V2.5-Pro

The distinction between 1.02T total parameters and 42B active parameters is important. MiMo-V2.5-Pro is an MoE model: each token activates only a subset of the available parameters. That substantially changes its inference compute profile compared with a dense 1T-parameter model, although the complete model still has enormous memory and deployment requirements.

What Are the Main Features of MiMo-V2.5-Pro?

1. Trillion-Scale Sparse MoE Architecture

MiMo-V2.5-Pro contains 1.02T total parameters and 42B activated parameters.

Its architecture contains 384 routed experts, with eight experts activated for each token. The model has 70 transformer layers, consisting of one dense layer and 69 MoE layers.

This provides substantially more representational capacity than the standard MiMo-V2.5 while avoiding the compute cost of activating all 1.02T parameters for every token.

The practical takeaway is:

MiMo-V2.5-Pro is a trillion-parameter model by total capacity, but its per-token computation is governed by its 42B active-parameter path.

2. 1 Million Token Context

The model supports up to 1M tokens of context.

This is one of its most important capabilities for autonomous agents because long-running workflows can accumulate:

• Tool outputs
• Source code
• Search results
• Documentation
• Intermediate reasoning state
• User instructions
• Execution logs

Instead of repeatedly summarizing and discarding old context, an agent can potentially retain a much larger working history.

Xiaomi's long-context evaluation specifically tests the model from 32K through 1M input tokens.

3. Hybrid Sliding-Window and Global Attention

MiMo-V2.5-Pro uses a 6:1 ratio of Sliding Window Attention (SWA) to Global Attention, with a 128-token sliding window.

The architecture contains:

• 60 SWA layers
• 10 global-attention layers
• 128-token SWA window

Xiaomi reports that this design reduces KV-cache storage by nearly 7× while preserving long-context performance through learnable attention-sink bias.

This is one of the most technically significant parts of the model.

A 1M context window is expensive if every layer performs full attention over the entire sequence. Hybrid attention reduces the amount of information each layer needs to attend to globally while retaining dedicated global-attention layers for long-range relationships.

4. Multi-Token Prediction

MiMo-V2.5-Pro contains three lightweight MTP modules.

Multi-Token Prediction allows the model to predict multiple future tokens in a manner that can be used for speculative-style decoding and inference acceleration.

Xiaomi reports that its MTP architecture can triple output speed during inference under its described deployment configuration.

This matters particularly for agents because an agent may generate large amounts of intermediate output over a long trajectory. Improving token generation speed can therefore have a meaningful impact on total task latency.

5. Agentic Reinforcement Learning

MiMo-V2.5-Pro's post-training pipeline includes:

• Supervised fine-tuning
• Large-scale agentic reinforcement learning
• Multi-Teacher On-Policy Distillation (MOPD)

The training objective is explicitly oriented toward complex agent behavior rather than only static benchmark question answering.

This is why the model's strongest evaluations are concentrated around coding, terminal interaction, long-context reasoning, and agent benchmarks.

6. 27T-Token Pretraining

MiMo-V2.5-Pro was pretrained on approximately 27 trillion tokens, using FP8 mixed precision and a native 32K sequence length before supporting the extended 1M context window.

The scale of pretraining, combined with the trillion-parameter MoE architecture, places MiMo-V2.5-Pro firmly in the current frontier-model class.

---

How Does MiMo-V2.5-Pro Perform on Benchmarks?

The published evaluation results are particularly useful because they include both general reasoning benchmarks and practical coding/agent evaluations.

The results show a particularly strong profile in software engineering and mathematical reasoning. The reported 78.9 on SWE-Bench Verified and 68.4 on Terminal-Bench 2.0 are especially relevant for developers building autonomous coding systems.

Long-Context Performance

The long-context results are arguably more interesting than the standard benchmarks.

On the GraphWalks evaluation, MiMo-V2.5-Pro maintains measurable performance at extreme context lengths:

Xiaomi reports that the previous MiMo-V2-Pro deteriorates rapidly beyond 128K and reaches 0.00 at 1M on both subtasks, whereas V2.5-Pro retains non-zero performance at the full 1M context.

This is a meaningful distinction: MiMo-V2.5-Pro's 1M context is not merely a theoretical maximum; the published long-context evaluation demonstrates useful performance deep into that window.

Real-World Agent Examples

These are Xiaomi's reported demonstrations rather than standardized benchmark scores.

MiMo-V2.5-Pro vs MiMo-V2.5

The two models share the same generation but target different workloads.

The choice is therefore not simply "Pro is better."

If an application needs native image, video, and audio understanding, MiMo-V2.5 is the more natural choice. If the workload centers on complex reasoning, software engineering, terminal interaction, and long-running agents, MiMo-V2.5-Pro is the stronger fit.

Limitations of MiMo-V2.5-Pro

1. Text-Only Model Input

Unlike mimo-v2.5, the Pro model's official specification lists Text as its input modality. It should not be marketed as the multimodal model in the V2.5 family.

2. High Compute Requirements for Self-Hosting

The model has approximately 1T total parameters / 42B active parameters, making local deployment considerably more demanding than conventional small open models.

3. Agent Performance Depends on the Framework

Xiaomi's near-1,000-tool-call claim is explicitly associated with using an appropriate Agent framework. The model alone does not guarantee that an application will achieve the same long-horizon performance.

4. Reasoning Content Handling

Multi-turn Agent applications using thinking mode and tool calls need to correctly preserve reasoning_content. Incorrect handling can produce API errors.

Best Use Cases

Large-scale software engineering

• Repository migration
• Refactoring
• Debugging
• Test generation
• Compiler development
• Full-stack application development

Autonomous coding Agents

MiMo-V2.5-Pro is especially well suited to Agents that need to repeatedly:

inspect → modify → execute → test → diagnose → modify

Long-document reasoning

Its 1M context makes it useful for:

• Legal contracts
• Technical specifications
• Large research collections
• Enterprise documentation

Complex business Agents

Tool calling + web search + structured output can support:

• Research Agents
• Data-processing Agents
• Enterprise workflow automation
• Multi-step decision systems

How to Use MiMo-V2.5-Pro API on CometAPI

The relevant CometAPI model ID is:

mimo-v2.5-pro

For developers, an API aggregation layer is particularly useful for testing MiMo-V2.5-Pro against other high-end reasoning and agent models without maintaining independent provider integrations.

Step 1: Get a CometAPI API Key

Create or sign into your CometAPI account and obtain your API key from the API console.

Set it as an environment variable:

export COMETAPI_KEY="YOUR_COMETAPI_KEY"

Step 2: Configure the API Client

Xiaomi provides an API compatible with both OpenAI and Anthropic protocols, making migration relatively straightforward. For production integration, use the current CometAPI endpoint/schema for mimo-v2.5-pro, particularly if you need advanced features such as tool calling, streaming, structured outputs, or long-context requests.

Step 3: Connect MiMo-V2.5-Pro to Tools

The model becomes substantially more useful when connected to external tools.

For example:

                    ┌── Web Search
                    │
User → MiMo-V2.5-Pro ├── GitHub
                    │
                    ├── Terminal
                    │
                    ├── Database
                    │
                    └── Internal APIs
                           ↓
                     Tool Results
                           ↓
                    MiMo-V2.5-Pro
                           ↓
                      Final Result

This architecture aligns much more closely with the model's intended use than a simple chatbot integration.