MiMo-V2.5
Xiaomi's open-weight, omnimodal AI with 1M context and strong agentic capabilities — built for developers who want sovereignty and uncensored, private inference.
Overview
What is MiMo-V2.5
MiMo-V2.5 is a 310B-parameter sparse Mixture-of-Experts (MoE) model from Xiaomi, released April 2026. It natively processes text, images, video, and audio within a unified architecture, supports up to 1 million tokens of context, and delivers powerful reasoning and tool-use capabilities for agentic workflows.
Running it privately on Venice
On Venice, MiMo-V2.5 runs with zero retention — your prompts are never stored or profiled. This open, MIT-licensed model can be used privately for uncensored, permissionless development, giving you full sovereignty over your AI workflows without Big Tech surveillance.
Assessment
Strengths and limitations
- Fully open weights under MIT license: ideal for self-hosting, fine-tuning, and commercial deployment without restrictions.
- Native omnimodal understanding: processes text, images, video, and audio in a single unified model.
- 1 million token context window enables ultra-long document reasoning and complex agentic workflows.
- Supports tool use, web search, and structured JSON output — optimized for autonomous agents and coding tasks.
- Cost-efficient at $0.14/$0.28 per million tokens (in/out), with cached input as low as $0.05/M.
- Slightly less intelligent than MiMo-V2.5-Pro and closed frontier models like Claude Opus on agentic benchmarks.
- Not uncensored: content moderation policies apply per Xiaomi's documentation.
- No end-to-end encryption or TEE execution on Venice, limiting privacy to zero retention only.
Capabilities
What it supports
- Tool use / function calling
- Vision (image input)
- Reasoning
- Web search
- Code-optimized
- Structured output (JSON schema)
- Audio input
- Video input
- Multiple image inputs
- Log probabilities
Specifications
Datasheet
- Maker
- Xiaomi
- Released
- April 22, 2026
- Architecture
- Sparse Mixture-of-Experts (MoE)
- Parameters
- 310B total, ~15B active per token
- Open weights
- Yes — MIT license
- Context window
- 1,000K tokens
- Max output
- 65.536K tokens
- Capabilities
- Vision, Function calling, Reasoning, Web search, Code-optimized
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Jun 2026
- License
- MIT
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "xiaomi-mimo-v2-5",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.14 per 1M input tokens and $0.28 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Max output | Open weights | Price (in/out per 1M) | Strongest at |
|---|---|---|---|---|
| MiMo-V2.5 | 65.536K | Yes | $0.14 in · $0.28 out / 1M | Omnimodal agents, open deployment |
| Claude Opus 5 | 65.536K | No | $6 in · $30 out / 1M | High-intelligence agentic tasks |
| DeepSeek V4 Flash 0731 | 65.536K | No | $0.17 in · $0.35 out / 1M | Speed and coding |
| Google Gemma 4 31B Instruct | 65.536K | Yes | $0.12 in · $0.36 out / 1M | Efficient small-scale reasoning |
Open, MIT-licensed model with native vision, audio, and video support — ideal for private, sovereign AI.
Use cases
What it is good for
- 01Building autonomous agents that process multimodal inputs and call external tools.
- 02Long-context analysis of legal, financial, or technical documents with image and text.
- 03Multimodal RAG pipelines using private, cached data with audio and visual content.
- 04Cost-sensitive production deployments where open weights and commercial use are required.
- 05Developers seeking full control over AI inference without vendor lock-in.
Prompting
Getting better results
Use structured JSON output for reliable parsing in agent pipelines.
Include image, audio, or video inputs directly — the model natively understands them.
Leverage web search and function calling for real-time data and actions.
Break complex tasks into steps — the 1M context supports deep reasoning chains.
Version history
Predecessor model
Current — omnimodal, 1M context, MIT-licensed
FAQ
Frequently asked questions
MiMo-V2.5 is a 310B-parameter sparse MoE model from Xiaomi, released in April 2026. It natively supports text, image, audio, and video understanding, with a 1 million token context window and strong agentic capabilities for autonomous workflows.
MiMo-V2.5 is not free, but it is open-weight and commercially usable under the MIT license. On Venice, you pay per token — $0.14 per million input tokens and $0.28 per million output tokens, with lower rates for cached input.
Yes, MiMo-V2.5 is fully open weights under the MIT license, meaning you can inspect, self-host, fine-tune, and deploy it commercially without restrictions.
Yes, MiMo-V2.5 natively supports image, video, and audio input through dedicated encoders, enabling true multimodal reasoning in a single model.
MiMo-V2.5 supports up to 1 million tokens of context, allowing it to process extremely long documents, transcripts, or complex reasoning chains in a single session.
Yes, MiMo-V2.5 supports tool use, web search, and function calling, making it well-suited for building autonomous agents and interactive applications.
MiMo-V2.5 is open and more affordable, with native multimodal support, while Claude Opus 5 has higher reasoning intelligence but is closed, more expensive, and lacks vision. Choose MiMo-V2.5 for sovereignty and multimodal use, Opus for maximum reasoning power.
Yes, on Venice, MiMo-V2.5 runs with zero retention — your prompts are never stored, profiled, or used for training, ensuring private, uncensored, and permissionless AI use.
Run MiMo-V2.5 privately
No prompt logging. No data used for training.