Xiaomi MiMo —
Reasoning-First, Open-Source LLM Family
A reasoning-first model family purpose-built for AI agents — optimized for complex reasoning, production coding, long-context tasks, and tool use. MIT licensed, from 7B edge to 1T+ MoE flagship. Made by Xiaomi.
Why MiMo?
MiMo tackles the hardest problems in production AI — high cost, slow inference, weak long-context — while keeping the whole family open under MIT.
MiMo Model Family
A layered lineup that balances reasoning capability with deployment flexibility — from a 7B edge model to a 1T+ MoE flagship.
MiMo-7B
7B reasoning-first, math and code strong, optimized for on-device. Ideal for offline, edge, or low-latency deployments in the Xiaomi ecosystem.
MiMo-V2-Flash (MoE)
309B total / 15B active parameters. 56k context at 150 tok/s with 73.4% SWE-Bench. The MoE flagship for long-context reasoning and high-throughput coding.
MiMo-V2.5-Pro NEW
1T+ total parameters with a 1M-token context window. Agent-optimized flagship for complex multi-step tasks. MIT licensed, available via API.
MiMo-V2.5-Omni NEW
Full-modal model handling image, video, audio, and text — perception and reasoning in a single pass. MIT licensed.
MiMo-V2.5-TTS NEW
Speech synthesis model with bilingual (EN/CN) and dialect support. Natural prosody and real-time performance for voice applications.
MiMo Code NEW
Terminal-native AI coding agent — open source under MIT. Infinite context, persistent memory, multi-agent switching. Integrates with VS Code, Cursor, Cline, Zed.
MiMo-VL
Vision-language model for perception-rich tasks — scene understanding, document parsing, and visual QA.
MiMo-Audio
Speech understanding and generation tuned for low-latency voice assistants. Optimized for wake-word accuracy and real-time interaction.
MiMo-Embodied
Cross-domain embodied model for robotics and autonomous driving — perception, planning, and control with safety-aware reasoning.
Architecture & Methods
MiMo's technical spine combines MoE, Hybrid Attention, and compression techniques to keep reasoning sharp while enabling edge deployment.
Deployment: Edge, Cloud, Hybrid
MiMo is engineered to meet you where you are — on-device for privacy and latency, cloud for scale, or hybrid for the best of both.
Use Cases — Human · Car · Home
MiMo powers end-to-end experiences across personal devices, vehicles, and smart homes — delivering proactive, context-aware intelligence.
Human (Mobile)
HyperOS integration delivers faster on-device responses, robust long-text handling, code reasoning, and multi-turn instructions — with offline fallback for weak-signal areas.
Car
In-cabin assistants handle navigation, entertainment, vehicle control, and safety-aware reasoning — edge-optimized for real-time voice and multi-turn dialogue.
Home
Xiaomi MiMo-powered smart home orchestration: proactive scene automation, IoT device control, and natural language interaction across the Xiaomi ecosystem.
Benchmarks & Proof
MiMo closes the gap with frontier closed-source models while staying deployable on everyday hardware — at 5-20% of their cost.
For context: GPT-5.5 scores ~88.7% on SWE-Bench Verified, Claude Opus 4.5 scores 76.80%. MiMo-V2-Flash's 73.4% puts it ahead of most models at a fraction of the cost.
Roadmap & Milestones
MiMo is iterating fast — from a single 7B reasoning model in early 2025 to a full multimodal, audio, and embodied lineup.
Developer Resources
Everything you need to adopt MiMo: code, weights, guides, and integrations.
curl -fsSL https://mimo.xiaomi.com/install | bash📦 Install via pip
pip install transformers torch accelerate
# Or clone from source
git clone https://github.com/XiaomiMiMo/mimo-models.git
cd mimo-models⚡ Quick Inference
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"XiaomiMiMo/MiMo-7B-Instruct",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(
"XiaomiMiMo/MiMo-7B-Instruct"
)
prompt = "Explain the MoE architecture in simple terms"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=256)
print(tokenizer.decode(outputs[0]))FAQ
Is MiMo open source?
Yes. Every MiMo model is released under the MIT License — weights, code, and tooling. Use it commercially, fine-tune it, distribute it. No strings attached.
How does pricing compare to GPT-5.5 or Claude Opus 4.5?
MiMo-V2.5-Pro costs $1/M input + $3/M output. GPT-5.5 is $5/$30, Claude Opus 4.5 is $15/$75. That's 80–95% cheaper on input, 80–96% cheaper on output. Open weights means zero per-token cost if you self-host.
What is MiMo Code?
MiMo Code is Xiaomi's first terminal-native AI coding agent (Coding Agent), released June 2026. MIT licensed. Features infinite context, persistent memory across sessions, and multi-agent switching. Works with VS Code, Cursor, Cline, Zed. Read more →
Can I run MiMo on my laptop?
MiMo-7B runs on consumer hardware with 8GB+ RAM. INT4 quantized versions go even smaller. V2-Flash and V2.5-Pro are best accessed via API or deployed on GPU servers.
How do I access the MiMo API?
Visit platform.xiaomimimo.com to sign up. The API is OpenAI-compatible — just swap the base URL. API guide →
Where can I download the models?
All models are on Hugging Face and GitHub under the MIT License. MiMo Code installs via curl -fsSL https://mimo.xiaomi.com/install | bash.