Xiaomi dropped the full MiMo V2.5 series on April 22-23, 2026. It is not one model. It is four: V2.5 Pro, V2.5 Standard, V2.5 TTS, and V2.5 ASR. Each one targets a different use case, from agentic coding to multimodal reasoning to voice synthesis and speech recognition.
This is the most complete product lineup Xiaomi has shipped for its AI platform. Previous releases focused on individual models. The V2.5 series is a coordinated family designed to cover the full stack of language, vision, and voice.
This guide breaks down every model in the family, compares them side by side, and helps you pick the right one for your project.
Quick comparison
| Feature | V2.5 Pro | V2.5 Standard | V2.5 TTS | V2.5 ASR |
|---|---|---|---|---|
| Primary use | Agentic coding | Multimodal reasoning | Text-to-speech | Speech recognition |
| Parameters | 1T+ (MoE) / 42B active | Not disclosed | Not disclosed | Not disclosed |
| Context window | 1M tokens | Large (not confirmed) | N/A | N/A |
| SWE-bench Pro | 57.2% | N/A | N/A | N/A |
| Modalities | Text, code | Image, audio, video, text | Text to audio | Audio to text |
| Speed | Baseline | Faster than Pro | Fast | Fast |
| Relative API cost | Higher | ~50% cheaper than Pro | Low | Low |
All four models are available through the Xiaomi API platform. Open-source releases are planned (more on that below).
MiMo V2.5 Pro
V2.5 Pro is the flagship. It is a mixture-of-experts model with over 1 trillion total parameters and 42 billion active parameters per forward pass. The context window stretches to 1 million tokens, which means it can ingest entire codebases in a single prompt.
The headline number is 57.2% on SWE-bench Pro, which puts it at the top of the leaderboard for agentic coding tasks. This is the model you want when you are building AI coding agents, automated PR reviewers, or any pipeline that needs to read, understand, and modify large repositories. For a deeper look at architecture, benchmarks, and practical usage, see the MiMo V2.5 Pro guide.
V2.5 Pro also performs well on general reasoning benchmarks, but its real strength is sustained, multi-step tool use. It can plan a sequence of file edits, execute them, and verify the results. If you used MiMo V2 Pro before, V2.5 Pro is a significant jump in both accuracy and context handling. For a comparison with other frontier models, check MiMo V2 Pro vs Claude vs GPT.
MiMo V2.5 Standard
V2.5 Standard is the multimodal workhorse. It natively processes images, audio, and video alongside text. You do not need separate preprocessing pipelines or adapters. Feed it a screenshot, a voice clip, or a short video and it reasons over all of them in a single call.
Performance is strong. Xiaomi claims V2.5 Standard outperforms the older V2 Pro on agent benchmarks, which is notable because V2 Pro was already competitive with models from OpenAI and Anthropic. The speed advantage matters too: V2.5 Standard returns responses faster than V2.5 Pro for most tasks, making it better suited for real-time applications and user-facing products.
The cost story is equally compelling. API pricing for V2.5 Standard runs roughly 50% cheaper than V2.5 Pro. If your application does not specifically need million-token code context or top-tier SWE-bench scores, Standard gives you more capability per dollar. It is the right default for most teams building multimodal apps, chatbots, or content analysis tools.
Practical examples: feed V2.5 Standard a product photo and ask it to write a listing description. Send it a meeting recording and get a structured summary. Upload a dashboard screenshot and ask it to identify anomalies. These workflows run faster and cheaper than routing through V2.5 Pro.
If you want to understand how it relates to earlier multimodal work from Xiaomi, see What is MiMo V2 Omni.
MiMo V2.5 TTS
V2.5 TTS handles text-to-speech synthesis. It generates natural-sounding audio from text input, supporting multiple languages and voice styles. Xiaomi has invested heavily in prosody and intonation, so the output sounds less robotic than many competing TTS systems.
The model supports voice cloning and style transfer, letting you create custom voice profiles for branded experiences. You can control speaking rate, pitch, and emotional tone through API parameters. This level of control makes it viable for production use cases where generic robot voices would hurt user experience.
Use cases include voice assistants, audiobook generation, accessibility features, and any product where you need to convert written content into spoken audio at scale. The model integrates directly with the Xiaomi API, so you can chain it with V2.5 Standard or V2.5 Pro outputs to build end-to-end pipelines that reason over data and then speak the results.
Latency is low enough for real-time streaming. You can start playing audio before the full response is generated, which is critical for conversational interfaces where users expect immediate feedback.
MiMo V2.5 ASR
V2.5 ASR is the speech recognition model. It converts spoken audio into text with high accuracy across multiple languages and accents. It handles noisy environments, overlapping speakers, and varied recording quality better than the previous generation.
The model supports both batch processing and real-time streaming transcription. For batch jobs, you upload audio files and get back structured transcripts with timestamps and speaker labels. For streaming, you send audio chunks and receive partial transcripts as the speaker talks. Both modes are accessible through the same API.
ASR accuracy is particularly strong for Mandarin and English, with solid support for other major languages. Xiaomi trained the model on a diverse dataset that includes phone calls, meetings, podcasts, and field recordings, so it generalizes well to real-world audio conditions.
Pair it with V2.5 TTS for full duplex voice interactions, or feed its transcriptions into V2.5 Standard or V2.5 Pro for downstream reasoning. This makes it possible to build complete voice-driven AI workflows entirely within the MiMo ecosystem. For context on how Xiaomi’s lighter models fit into this stack, see What is MiMo V2 Flash.
How to choose the right model
Before diving into the decision tree, it helps to understand that these models are designed to work together. You are not locked into picking one. Many production systems will use two or three V2.5 models in combination.
The decision tree is straightforward:
- Building AI coding agents or code automation? Use V2.5 Pro. Nothing else in the family matches its SWE-bench scores or million-token context.
- Building multimodal apps (image, audio, video understanding)? Use V2.5 Standard. It is faster, cheaper, and handles multiple input types natively.
- Need voice output? Use V2.5 TTS for speech synthesis.
- Need voice input? Use V2.5 ASR for transcription.
- Building a full voice assistant? Combine V2.5 ASR + V2.5 Standard (or Pro) + V2.5 TTS in a pipeline.
For most teams, V2.5 Standard is the starting point. Only move to V2.5 Pro if you specifically need top-tier code reasoning or the extended context window. The cost savings from Standard add up fast at scale.
Here is a quick reference:
| Your goal | Recommended model | Why |
|---|---|---|
| AI coding agent | V2.5 Pro | Best SWE-bench scores, 1M context |
| Multimodal chatbot | V2.5 Standard | Native image/audio/video, lower cost |
| Image analysis app | V2.5 Standard | Strong vision capabilities |
| Voice assistant | ASR + Standard + TTS | Full pipeline within one ecosystem |
| Transcription service | V2.5 ASR | Streaming and batch support |
| Audiobook generation | V2.5 TTS | Natural prosody, voice cloning |
| General reasoning | V2.5 Standard | Good performance at lower cost |
Token Plan pricing
Xiaomi updated its Token Plan pricing alongside the V2.5 launch. Key changes:
| Plan | Included tokens | V2.5 Pro rate | V2.5 Standard rate | Notes |
|---|---|---|---|---|
| Free tier | Limited daily quota | Standard pricing | Standard pricing | Good for testing |
| Basic | Moderate allocation | Discounted | Discounted | Suitable for small projects |
| Pro | Large allocation | Further discounted | Further discounted | Best per-token value |
| Enterprise | Custom | Custom | Custom | Volume agreements available |
V2.5 Standard tokens cost roughly half what V2.5 Pro tokens cost across all plans. TTS and ASR are billed separately based on audio duration rather than token count. Check Xiaomi’s API pricing page for exact current rates, as these may shift after launch.
Open-source plans
Xiaomi has confirmed that open-source releases are coming for the V2.5 series. The previous generation (V2 Pro, V2 Flash) was released under permissive licenses on Hugging Face and ModelScope, and the same approach is expected here.
No exact date has been announced. Based on past patterns, expect open weights within weeks to a few months after the API launch. The open-source versions will likely include V2.5 Pro and V2.5 Standard at minimum. TTS and ASR availability is less certain.
This matters if you need to self-host for data privacy, fine-tune on proprietary data, or run inference on your own hardware. Keep an eye on Xiaomi’s GitHub and Hugging Face pages for announcements.
When the weights drop, expect the community to produce quantized versions (GGUF, AWQ, GPTQ) quickly. The 42B active parameter count for V2.5 Pro means it should be runnable on high-end consumer GPUs with quantization, similar to how the community handled previous MoE models.
For broader context on open models from Chinese labs, see Best Chinese AI models 2026.
What changed from V2 to V2.5
The jump from V2 to V2.5 is not incremental. Here are the key differences:
- Unified family. V2 shipped individual models at different times. V2.5 launched as a coordinated product line with shared API conventions and designed interoperability.
- Agentic coding leap. V2 Pro was strong at code generation. V2.5 Pro is built specifically for agentic workflows: multi-step planning, tool use, and self-verification across million-token contexts.
- Native multimodal. V2 required separate models or adapters for different modalities. V2.5 Standard handles image, audio, and video in a single model with a single API call.
- Voice models. V2 had no dedicated TTS or ASR offerings. V2.5 adds both, making it possible to build complete voice pipelines without third-party services.
- Cost reduction. V2.5 Standard delivers better performance than V2 Pro at roughly half the API cost. Xiaomi clearly optimized for inference efficiency in this generation.
If you are currently using V2 Pro or V2 Flash, the migration path is straightforward. The API format is compatible, and most applications will see immediate improvements by switching to the V2.5 equivalents.
FAQ
Which MiMo V2.5 model is best for coding?
V2.5 Pro. It scores 57.2% on SWE-bench Pro, has 1 million token context, and is specifically optimized for agentic coding workflows. V2.5 Standard is capable at code tasks too, but Pro is the clear choice for serious code automation. See the full MiMo V2.5 Pro guide for details.
Is V2.5 Standard better than V2.5 Pro?
It depends on the task. V2.5 Standard is faster, cheaper, and handles multimodal inputs (images, audio, video) natively. It even outperforms the older V2 Pro on agent benchmarks. But V2.5 Pro still leads on pure coding tasks and offers the largest context window. Pick based on your use case, not a blanket “better” label.
Can I use V2.5 TTS and ASR together?
Yes. Chain V2.5 ASR (speech to text) with any reasoning model (Standard or Pro), then pass the output to V2.5 TTS (text to speech) for a complete voice interaction loop. The API supports this kind of pipeline natively.
When will MiMo V2.5 be open-source?
Xiaomi has confirmed open-source releases are planned but has not given an exact date. Based on previous releases, expect open weights within weeks to a few months. V2.5 Pro and V2.5 Standard are the most likely candidates for open release first.