Long-horizon coding agents
Run repository-scale work that may require many edits, tests, terminal calls, and recovery steps.
Independent tool overview
MiMo-V2.5-Pro at a glance
MiMo-V2.5-Pro is Xiaomi's MIT-licensed, text-only Mixture-of-Experts model for long-horizon coding and agent tasks, with 1.02 trillion total parameters, 42 billion active parameters, and a 1-million-token context window.
Visit the official MiMo-V2.5-Pro site ↗
- Model type
- Text-only agentic and coding MoE language model
- Parameters
- 1.02T total; 42B active
- Context window
- 1,048,576 tokens
- Maximum output
- 131,072 tokens in Xiaomi's integration guide
- License
- MIT
- Access
- API, Token Plan, AI Studio, or self-hosted weights
Overview
What MiMo-V2.5-Pro is
MiMo-V2.5-Pro is Xiaomi MiMo's flagship open-weight language model for demanding software engineering, reasoning, and tool-using agents. It is available through Xiaomi's pay-as-you-go API, Token Plan subscriptions, AI Studio, and downloadable FP8 weights.
The model uses a 1.02-trillion-parameter Mixture-of-Experts architecture while activating 42 billion parameters per token. Hybrid sliding-window and global attention reduce long-context cache pressure, and three Multi-Token Prediction modules are designed to increase decoding throughput.
This Pro variant is text-only despite its 1-million-token context. Xiaomi's separate MiMo-V2.5 model supports text and image input. Teams integrating Pro into coding agents also need a compatible harness that preserves the model's reasoning content across tool-call turns.
Use cases
Who MiMo-V2.5-Pro is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Large-codebase analysis
Provide extensive source, documentation, and task history in one long context.
Tool-driven engineering
Pair the model with a capable harness for search, file editing, commands, testing, and structured tools.
Cost-sensitive API workloads
Use low published token prices for high-volume text reasoning and code generation.
Controlled model deployment
Run MIT-licensed weights on owned infrastructure when the organization can support a trillion-parameter MoE.
Capabilities
Core MiMo-V2.5-Pro features
Million-token context
Supports up to 1,048,576 input tokens for large repositories and long-running agent histories.
Long agent trajectories
Post-training targets coherent tool use and instruction following across extended, multi-step tasks.
Agentic coding
Designed for repository understanding, implementation, review, structured artifacts, and software-engineering workflows.
Hybrid attention
Interleaves local sliding-window and global attention at a 6:1 ratio to reduce KV-cache requirements.
Multi-Token Prediction
Uses three native MTP modules intended to improve inference throughput.
OpenAI-compatible API
Xiaomi provides a compatible endpoint for pay-as-you-go access and region-specific Token Plans.
Open weights
Publishes instruct and base checkpoints under the permissive MIT license.
Self-hosting support
Provides official deployment guidance for SGLang and vLLM, including reasoning and tool-call parsers.
Process
How the MiMo-V2.5-Pro workflow works
Step 1
Choose hosted or self-hosted
Compare data requirements and sustained volume against the infrastructure needed for 1.02T FP8 weights.
Step 2
Select an agent harness
Use a coding agent that supports Xiaomi's endpoint, thinking mode, long context, and tool-call replay rules.
Step 3
Configure the model contract
Set the correct model ID, context, output cap, parser, and provider-specific reasoning fields.
Step 4
Give scoped tools
Restrict file, terminal, network, deployment, and credential access to what the task actually needs.
Step 5
Start with a bounded task
Test repository navigation, edits, tests, and recovery on a representative job before long autonomous runs.
Step 6
Preserve tool-call history
Replay the actual reasoning_content and assistant content required by the API across tool turns.
Step 7
Measure the full trajectory
Track pass rate, tokens, wall time, retries, tool errors, regressions, and human review—not benchmark scores alone.
Cost
MiMo-V2.5-Pro pricing and free plan
Xiaomi offers very low pay-as-you-go token prices plus fixed Token Plan subscriptions. Pay-as-you-go separates cached input, uncached input, and output. Token Plans use dedicated regional endpoints and quota-based access; confirm the current included Credits in Subscription Management.
Pay-as-you-go API
$0.0036 cached input / $0.435 uncached input / $0.87 output per 1M tokens
Usage-based access through Xiaomi's OpenAI-compatible API.
- Cached input is priced separately
- Suitable for light or variable use
- Model ID is mimo-v2.5-pro
- External tools and infrastructure can add cost
Lite Token Plan
$6/month or $5.28/month effective annually
Entry subscription for quota-based model access.
- Dedicated regional Base URL and API key
- Annual billing advertised at 12% savings
- Current quota shown in Subscription Management
Standard Token Plan
$16/month or $14.08/month effective annually
Mid-volume subscription for individual development workflows.
- Covers the V2.5 model series
- Dedicated regional endpoint
- Quota-based rather than unrestricted
Pro and Max Token Plans
$50 or $100 monthly; $44 or $88/month effective annually
Higher-quota subscriptions for heavier agent and coding use.
- Annual prices reflect the advertised 12% saving
- MiMo Claw included on the current plan page
- Review fair-use, concurrency, and quota details before purchase
Self-hosted weights
Free MIT license; infrastructure costs apply
Download and serve the FP8 model with compatible inference software.
- 1.02T total parameters
- Data-center-class multi-GPU deployment
- Storage, compute, networking, and operations are separate
- Quantized community variants are not Xiaomi's official FP8 checkpoint
Pricing checked . Check current pricing at the source ↗
Assessment
MiMo-V2.5-Pro strengths and limitations
Where it stands out
- One-million-token context is well suited to large repositories and long tool histories
- 42B active parameters make the trillion-parameter MoE more efficient than a dense model of the same total size
- Official API pricing is unusually low for a flagship agentic model
- MIT-licensed weights allow commercial modification and self-hosting
- Xiaomi publishes detailed architecture, deployment, and evaluation information
- OpenAI-compatible access works with a growing set of coding-agent harnesses
- Pay-as-you-go and fixed Token Plans serve different workload patterns
What to consider
- MiMo-V2.5-Pro accepts text only; use the separate MiMo-V2.5 model for image input
- A one-million-token limit does not guarantee accurate recall or reasoning across every part of a huge prompt
- The 1.02T FP8 checkpoint demands substantial storage and multi-GPU infrastructure to self-host
- Tool-call integrations must preserve real reasoning_content and assistant content across turns
- Provider compatibility varies; not every OpenAI-compatible client handles thinking and tool calls correctly
- Vendor demonstrations and benchmark results should be validated on the organization's own repositories and harness
- Long autonomous trajectories expand security, regression, and cost exposure even at low per-token prices
- The Pro checkpoint's text-only scope makes it unsuitable for visual UI, diagram, or screenshot tasks
Compare
MiMo-V2.5-Pro alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Coding
Devstral 2
Choose Devstral 2 for Mistral's coding-specific open model family and a different deployment-to-capability tradeoff.
Explore Devstral 2 →Coding
Kimi K2.7 Code
Evaluate Kimi K2.7 Code for another current open-weight model focused on coding agents and leaner reasoning.
Explore Kimi K2.7 Code →Business Operations
DeepSeek
Consider DeepSeek for a broader family of low-cost reasoning and coding models with its own API and open weights.
Explore DeepSeek →Consumer
Trinity-Large-Thinking
Use Trinity Large Thinking for another permissively licensed long-context agent model with different routing and hosting options.
Explore Trinity-Large-Thinking →Questions
MiMo-V2.5-Pro FAQs
What is MiMo-V2.5-Pro?
It is Xiaomi's flagship text-only open-weight MoE model for long-horizon coding, reasoning, and tool-using agents.
Is MiMo-V2.5-Pro open source?
Xiaomi publishes the weights, tokenizer, and model card under the permissive MIT license.
How large is MiMo-V2.5-Pro?
It has 1.02 trillion total parameters and activates 42 billion parameters per token. The official instruct checkpoint uses FP8 mixed precision.
What is its context window?
The instruct model supports up to 1,048,576 tokens. Xiaomi's current agent integration guide lists a 131,072-token maximum output.
Does MiMo-V2.5-Pro accept images?
No. The Pro integration is text-only. Xiaomi's separate MiMo-V2.5 model supports text and image input.
How much does the API cost?
Pay-as-you-go pricing is $0.0036 per million cached input tokens, $0.435 per million uncached input tokens, and $0.87 per million output tokens.
Can MiMo-V2.5-Pro be self-hosted?
Yes, with MIT-licensed weights and official SGLang and vLLM guidance, but the trillion-parameter FP8 model requires serious multi-GPU infrastructure.
What integration detail matters for tool use?
A compatible harness must preserve the model's actual reasoning_content and assistant content in conversation history across tool-call turns.
Bottom line
Our MiMo-V2.5-Pro verdict
MiMo-V2.5-Pro is compelling for text-only coding agents that need very long context and low hosted inference cost. The MIT license and detailed deployment material make it more open than many flagship alternatives. The practical decision is less about headline benchmarks and more about the harness: verify tool-call replay, long-context behavior, repository pass rate, and safe permissions. Use Xiaomi's API for an inexpensive proof before considering the formidable cost of self-hosting a 1.02T FP8 model.
Visit MiMo-V2.5-Pro website ↗Similar Consumer Tools
View all Consumer AI tools →
Nemotron 3 Nano Omni
Nemotron 3 Nano Omni - NVIDIA's new open model combining vision, audio, and text
.jpeg)
Grok 4.3
Grok 4.3 - xAI’s latest model release with strong cost efficiency, instruction following, and domain-specific performance
GPT 5.5
GPT-5.5 - OpenAI's new frontier flagship model for agentic coding and computer use
ERNIE 5.1
ERNIE 5.1 - Baidu's new foundation model that ranks No. 4 on Arena search and claims 94% lower training costs
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.