Xiaomi MiMo-V2-Pro: A 1T-Parameter Model for Coding and AI Agents
Xiaomi MiMo-V2-Pro is a trillion-parameter AI model that beats Claude Sonnet 4.6 at coding, approaches Claude Opus 4.6 on agent tasks, and costs 67% less. After processing 1T tokens on OpenRouter as “Hunter Alpha,” Xiaomi officially released it with a 1M-token context window and one week of free API access.
You probably know Xiaomi for smartphones. Then electric cars. Now the company is competing with Anthropic, OpenAI, and other foundation-model providers.
On March 18, 2026, Xiaomi’s MiMo team announced MiMo-V2-Pro, a foundation model designed for agentic workloads. Before the announcement, Xiaomi deployed it on OpenRouter under the codename “Hunter Alpha.” The model topped daily usage charts and processed more than 1T tokens before users knew who built it.
This guide covers the model’s benchmarks, pricing, architecture, API access, and integration patterns for coding and agent workflows.
MiMo-V2-Pro at a Glance
| Feature | MiMo-V2-Pro | Claude Sonnet 4.6 | Claude Opus 4.6 |
|---|---|---|---|
| Pricing (input/output) | $1/$3 per 1M tokens | $3/$15 per 1M tokens | $5/$25 per 1M tokens |
| Context window | 1M tokens | 200K tokens | 200K tokens |
| SWE-bench Verified | 78.0% | 79.6% | 80.8% |
| ClawEval | 61.5% | 66.3% | 66.3% |
| PinchBench | 84.0% | 86.9% | 86.3% |
| Architecture | 1T total, 42B active | ~200B estimated | ~400B estimated |
| Free access | One week on OpenRouter | Paid only | Paid only |
The short version: MiMo-V2-Pro is designed for agents that orchestrate multi-step workflows, use tools, and work across large codebases—not just for conversational question answering.
From Smartphones to Foundation Models
Xiaomi built its business around smartphones before expanding into electric vehicles with the SU7. MiMo-V2-Pro represents another strategic expansion: building foundation models for global AI workloads.
The company is not only adding AI features to its existing products. It is also competing directly on model quality, context length, agent performance, and inference cost.
The Hunter Alpha Deployment
One week before the official launch, an anonymous model called “Hunter Alpha” appeared on OpenRouter. Developers noticed that it:
- Topped OpenRouter’s daily usage charts on multiple days
- Processed more than 1T tokens during the stealth deployment
- Ranked #8 globally on the Artificial Analysis Intelligence Index
- Ranked #2 among Chinese LLMs, behind DeepSeek
Hunter Alpha was a beta deployment rather than a leak.
MiMo-V2-Pro is the production release. Xiaomi used the deployment period to collect user feedback and improve long-context handling and agent-scenario stability.
Performance: Coding and Agent Benchmarks
Xiaomi published benchmark comparisons against Claude and GPT models.
Agent Performance
| Benchmark | MiMo-V2-Pro | Claude Opus 4.6 | Claude Sonnet 4.6 | GPT-5.2 |
|---|---|---|---|---|
| ClawEval | 61.5% | 66.3% | 66.3% | 50.0% |
| PinchBench | 84.0% | 86.3% | 86.9% | 77.0% |
| GDPVal-AA | 96.8 | 99.3 | 97.9 | 98.7 |
| τ2-bench (Telecom) | 93.5 | 98.0 | 97.9 | 98.0 |
On these agent benchmarks, MiMo-V2-Pro sits between Sonnet and Opus. It does not match Opus across every test, but its results are close enough to make it relevant for developers building tool-using workflows.
Coding Performance
| Benchmark | MiMo-V2-Pro | Claude Opus 4.6 | Claude Sonnet 4.6 | GPT-5.2 |
|---|---|---|---|---|
| SWE-bench Verified | 78.0% | 80.8% | 79.6% | 80.0% |
| SWE-bench Multilingual | 71.7% | 77.8% | 75.9% | 72.0% |
| Terminal-Bench 2.0 | 57.1% | 65.4% | 59.1% | 54.0% |
| DeepSearch QA-F1 | 86.7% | 91.3% | 89.2% | 79.0% |
MiMo-V2-Pro scores higher than Claude Sonnet 4.6 on SWE-bench Verified, although it trails Sonnet on the other listed coding benchmarks.
Xiaomi’s internal engineers described the coding experience as approaching Claude Opus 4.6, particularly in system design, code style, and problem-solving efficiency.
Architecture and Context Length
MiMo-V2-Pro combines a large parameter count with a smaller active parameter set during inference:
- More than 1T total parameters
- 42B active parameters during inference
- Hybrid Attention with a 7:1 ratio, up from 5:1 in V2-Flash
- 1M-token context window
- MTP, or Multi-Token Prediction, for faster generation
- Three times larger than MiMo-V2-Flash
The 1M-token context window is useful for workflows that need to process large repositories, long documentation sets, or multi-step agent histories without aggressively summarizing earlier context.
Pricing
MiMo-V2-Pro uses different rates depending on the context length.
| Model | Input | Output | Cache read | Cache write |
|---|---|---|---|---|
| MiMo-V2-Pro, up to 256K | $1 per 1M | $3 per 1M | $0.20 | Free |
| MiMo-V2-Pro, 256K–1M | $2 per 1M | $6 per 1M | $0.40 | Free |
| Claude Sonnet 4.6 | $3 per 1M | $15 per 1M | $0.30 | $3.75 |
| Claude Opus 4.6 | $5 per 1M | $25 per 1M | $0.50 | $6.25 |
At the lower context tier, MiMo-V2-Pro costs one-third as much as Sonnet for input tokens and one-fifth as much for output tokens. Cache writes are free during the launch period.
For context-heavy workloads, the 256K–1M pricing tier still costs less than Claude Sonnet 4.6’s listed rates.
Free Access
Xiaomi is offering one week of free API access through five agent frameworks:
- OpenClaw
- OpenCode
- KiloCode
- Blackbox
- Cline
If your workflow uses one of these frameworks, the launch period is an opportunity to test the model against real tasks before paying for usage.
Why MiMo-V2-Pro Is Built for Agents
MiMo-V2-Pro targets workflows that involve:
- Planning a task
- Calling tools
- Reading tool output
- Updating the plan
- Editing files or making API calls
- Running tests
- Iterating based on results
That makes it a better fit for software agents than a workflow that only sends a single prompt and displays a response.
OpenClaw Integration
MiMo-V2-Pro was fine-tuned for OpenClaw using supervised fine-tuning and reinforcement learning on complex agent scaffolds. The integration focuses on:
- Tool-call stability across multi-step workflows
- Large context windows for high-intensity Claw applications
- Compatibility with OpenClaw’s orchestration layer
On OpenClaw’s standard benchmarks, MiMo-V2-Pro ranks third globally:
- PinchBench: 84.0%
- ClawEval: 61.5%
Example Agent Tasks
During the Hunter Alpha deployment, developers used the model for:
- Multi-file code generation across repositories
- API debugging with iterative testing
- Database schema design and migration generation
- Frontend component generation with styling
Coding tools were the largest usage category by volume during the test.
Frontend Generation Example
Xiaomi shared examples of MiMo-V2-Pro generating complete frontend applications through OpenClaw.
1990s Magazine-Style Website
The prompt specified a complete visual direction:
Mimic 1990s print magazine aesthetics. Title in serif font like Playfair Display,
body in monospace like IBM Plex Mono. Magazine-style multi-column grid with uneven
column widths. Large titles offset left beyond the viewport to suggest print bleed.
Images with sepia(0.2) filter and noise overlay. Page transitions mimicking page-turn
effects. Navigation styled as a magazine table of contents, each item numbered
01/02/03, numbers enlarge on hover. Footer designed as a magazine colophon with a
fake ISSN number. Paper texture background.
The model generated a functional website matching the requested specifications in a single pass.
How to Access the MiMo-V2-Pro API
You can access the model through OpenRouter, Xiaomi’s direct API, or a supported agent framework.
Option 1: OpenRouter
OpenRouter provides access to MiMo-V2-Pro during the one-week free launch period.
Step 1: Create an OpenRouter API key
- Create an account at
openrouter.ai. - Sign up with GitHub or email.
- Open the API Keys section.
- Generate a new key.
- Store the key in an environment variable rather than committing it to source control.
export OPENROUTER_API_KEY="your-openrouter-key"
Step 2: Send a request
import json
import os
import requests
url = "https://openrouter.ai/api/v1/chat/completions"
headers = {
"Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}",
"Content-Type": "application/json",
}
payload = {
"model": "xiaomi/mimo-v2-pro",
"messages": [
{
"role": "user",
"content": "Build a REST API with user authentication in Python",
}
],
}
response = requests.post(url, headers=headers, json=payload)
response.raise_for_status()
print(json.dumps(response.json(), indent=2))
Step 3: Inspect requests with Apidog
Agent responses can contain large JSON payloads and multiple tool-call messages. An API client such as Apidog can help you:
- Inspect complete JSON responses with syntax highlighting
- Trace multi-turn conversations
- Test parameters such as
temperatureandmax_tokens - Share debugging sessions with your team
Import the OpenRouter API specification into Apidog and use it to send and debug requests.
Option 2: Direct Xiaomi API
For production use, you can access MiMo-V2-Pro directly from Xiaomi.
Step 1: Create API credentials
- Visit
platform.xiaomimimo.com. - Create an account.
- Open API management.
- Generate credentials.
- Store the API key securely.
Step 2: Configure the client
import os
import requests
api_key = os.environ["XIAOMI_API_KEY"]
endpoint = "https://api.xiaomimimo.com/v1/chat/completions"
headers = {
"Authorization": f"Bearer {api_key}",
"Content-Type": "application/json",
}
payload = {
"model": "mimo-v2-pro",
"messages": [
{
"role": "system",
"content": "You are a helpful coding assistant.",
},
{
"role": "user",
"content": "Create a FastAPI endpoint with JWT authentication",
},
],
"temperature": 0.7,
"max_tokens": 8192,
}
response = requests.post(endpoint, headers=headers, json=payload)
response.raise_for_status()
print(response.json())
Option 3: Agent Frameworks
If you use OpenClaw, OpenCode, KiloCode, Blackbox, or Cline, check the framework documentation for MiMo-V2-Pro configuration. Free access is available for one week.
An OpenClaw configuration can look like this:
from openclaw import Agent
agent = Agent(
model="xiaomi/mimo-v2-pro",
tools=["file_system", "terminal", "browser"],
context_limit=1_000_000,
)
result = agent.run("Analyze this codebase and suggest improvements")
print(result)
MiMo-V2-Pro Compared with Other Models
MiMo-V2-Pro vs. Claude Sonnet 4.6
| Aspect | MiMo-V2-Pro | Claude Sonnet 4.6 |
|---|---|---|
| Pricing | $1/$3 per 1M | $3/$15 per 1M |
| Context | 1M tokens | 200K tokens |
| SWE-bench Verified | 78.0% | 79.6% |
| ClawEval | 61.5% | 66.3% |
| PinchBench | 84.0% | 86.9% |
| Free tier | One week | No |
Choose MiMo-V2-Pro when you need a larger context window, lower listed token costs, or want to test the model during the free week.
Choose Sonnet 4.6 when your team already uses the Anthropic ecosystem and needs the higher scores shown on these agent benchmarks.
MiMo-V2-Pro vs. Claude Opus 4.6
| Aspect | MiMo-V2-Pro | Claude Opus 4.6 |
|---|---|---|
| Pricing | $1/$3 per 1M | $5/$25 per 1M |
| Context | 1M tokens | 200K tokens |
| SWE-bench Verified | 78.0% | 80.8% |
| ClawEval | 61.5% | 66.3% |
| PinchBench | 84.0% | 86.3% |
Choose MiMo-V2-Pro when cost and context length are more important than maximizing benchmark performance.
Choose Opus 4.6 when you need the highest performance on the listed agent benchmarks and cost is less important.
MiMo-V2-Pro vs. GPT-5.2
| Aspect | MiMo-V2-Pro | GPT-5.2 |
|---|---|---|
| Pricing | $1/$3 per 1M | Varies |
| Context | 1M tokens | 128K tokens |
| SWE-bench Verified | 78.0% | 80.0% |
| ClawEval | 61.5% | 50.0% |
| PinchBench | 84.0% | 77.0% |
Choose MiMo-V2-Pro when you need the listed agent performance and a longer context window.
Choose GPT-5.2 when your team is already invested in the OpenAI ecosystem.
Practical Use Cases
1. API Development and Testing
Use an agent with filesystem, package-manager, and test-runner tools to generate and validate an API project:
api_agent = Agent(
model="xiaomi/mimo-v2-pro",
tools=["file_system", "package_manager", "test_runner"],
)
result = api_agent.run("""
Create a FastAPI application with:
- JWT authentication
- User registration and login endpoints
- Protected resource routes
- Pytest test suite
- Docker configuration
""")
The important implementation detail is the test runner. Asking the model to generate code is different from allowing it to execute tests and iterate on failures.
2. Legacy Code Migration
The 1M-token context window can be used to analyze large repositories or migration inputs:
migration_agent = Agent(
model="xiaomi/mimo-v2-pro",
context_window=1_000_000,
)
legacy_code = load_repository("./legacy-python-2")
result = migration_agent.run("""
Analyze this Python 2 codebase and:
1. Identify all Python 2-specific syntax
2. Generate Python 3.11-compatible versions
3. Create a migration report
4. Suggest modernization improvements
""")
For a production migration, split the work into reviewable changes and run the generated code through your existing test and deployment pipelines.
3. Multi-File Project Scaffolding
You can also use the model to generate a complete project structure:
scaffold_agent = Agent(
model="xiaomi/mimo-v2-pro",
tools=["file_system"],
)
result = scaffold_agent.run("""
Create a full-stack Next.js application with:
- App Router structure
- TypeScript configuration
- Tailwind CSS setup
- Authentication with NextAuth
- Database integration with Prisma
- API routes for CRUD operations
- Complete test suite with Jest
""")
Review generated dependencies, authentication flows, database migrations, and test coverage before using the scaffold in production.
Limitations and Considerations
MiMo-V2-Pro is promising, but it has tradeoffs.
Known Limitations
- Ecosystem maturity: The community is smaller than the OpenAI and Anthropic communities.
- Documentation gaps: Some features may not have detailed guides.
- Tool integration: There are fewer pre-built integrations than with established providers.
- Enterprise support: Production support and SLA options may be more limited.
When Not to Use MiMo-V2-Pro
Consider another provider when:
- You need enterprise-grade support and formal SLAs.
- Your team depends heavily on existing Anthropic or OpenAI integrations.
- You require guaranteed uptime commitments.
- You are building a consumer-facing product that needs extensive safety tuning.
Conclusion
MiMo-V2-Pro is worth testing if you:
- Need coding performance close to leading proprietary models
- Want to reduce token costs
- Need a 1M-token context window
- Are building agent workflows with OpenClaw or similar frameworks
- Want to evaluate the model during the free OpenRouter period
Test it against your actual workload rather than relying only on benchmark scores. Start with a representative repository, API workflow, or tool-calling task, then compare output quality, latency, failure recovery, and total cost with your current model.
Choose another provider if you need mature enterprise support, formal SLAs, extensive pre-built integrations, or specialized safety tuning.
FAQ
Is MiMo-V2-Pro free on OpenRouter?
Yes, for one week as part of the launch promotion. After the free period, the listed standard pricing is $1/$3 per 1M tokens for contexts up to 256K.
How does MiMo-V2-Pro compare with Claude Sonnet 4.6?
MiMo-V2-Pro scores 78.0% on SWE-bench Verified compared with Sonnet’s 79.6%. On ClawEval, it scores 61.5% compared with Sonnet’s 66.3%.
The tradeoff is lower listed pricing and a larger context window: 1M tokens compared with 200K.
Can I use MiMo-V2-Pro for commercial projects?
The API terms allow commercial use. Check Xiaomi’s platform documentation for the licensing details that apply to your project.
What is the maximum context length?
MiMo-V2-Pro supports up to 1M tokens. The listed pricing tiers are:
- Up to 256K tokens: $1 input and $3 output per 1M tokens
- 256K to 1M tokens: $2 input and $6 output per 1M tokens
Does MiMo-V2-Pro support function calling?
Yes. The model is optimized for tool use and agent workflows. Its listed benchmark results include 84.0% on PinchBench and 61.5% on ClawEval.
How do I get API access after the free week?
You can use:
- OpenRouter with standard paid pricing
- The direct Xiaomi API at
platform.xiaomimimo.com - Agent frameworks including OpenClaw, OpenCode, KiloCode, Blackbox, and Cline
Is there a self-hosted option?
No. MiMo-V2-Pro is currently available through APIs. Xiaomi has not announced a self-hosted deployment option.
Next Steps
- Test the API: Create an account on OpenRouter or the Xiaomi platform.
- Send a representative request: Use a real coding, debugging, or agent workflow.
- Inspect the responses: Import the API specification into Apidog to debug requests and responses.
- Try an agent framework: Evaluate the OpenClaw integration or another supported framework.
- Compare costs and quality: Measure token usage, latency, tool-call reliability, and task completion.
Top comments (0)