DEV.co
Open-Source LLM · XiaomiMiMo

MiMo-V2.5-Pro

MiMo-V2.5-Pro is a 1.02T-parameter Mixture-of-Experts language model from Xiaomi with 42B active parameters, designed for long-context reasoning (up to 1M tokens) and agentic tasks. It uses a hybrid attention architecture combining sliding-window and global attention, plus multi-token prediction for faster inference. The model is MIT-licensed, ungated, and available on HuggingFace. It shows strong performance on math, code, and long-context benchmarks, particularly excelling at complex reasoning tasks and software engineering problems.

Source: HuggingFace — huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro
1023.2B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
101.7k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperXiaomiMiMo
Parameters1023.2B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads101.7k
Likes698
Last updated2026-05-08
SourceXiaomiMiMo/MiMo-V2.5-Pro

What MiMo-V2.5-Pro is

MiMo-V2.5-Pro is a MoE-based decoder-only language model with 70 layers (1 dense + 69 MoE), 384 routed experts (8 per token), hidden size 6144, 128 attention heads, and GQA. Key architectural innovations: (1) Hybrid attention—60 sliding-window layers (128-token window) + 10 global attention layers in 6:1 ratio, reducing KV-cache by ~7x; (2) Three native multi-token prediction modules using dense FFNs for 3x output speed during inference; (3) FP8 (E4M3) mixed precision training on 27T tokens with native 32k sequence length, supporting up to 1M via position interpolation. Post-training uses three-stage approach: SFT, domain-specialized RL with expert teachers, and multi-teacher on-policy distillation (MOPD). Evaluation shows competitive performance on MMLU (89.4), GSM8K (99.6), and strong long-context retention (0.37 BFS / 0.62 Parents at 1M tokens on GraphWalks).

Quickstart

Run MiMo-V2.5-Pro locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="XiaomiMiMo/MiMo-V2.5-Pro")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Agent & Tool-Use Workflows

Designed explicitly for sustained agentic execution. Post-training optimized for thousands of tool calls with strong coherence. Strong on complex reasoning and instruction-following makes it suitable for autonomous systems requiring multi-step planning and external integration.

Long-Context Document & Code Analysis

1M-token context window with hybrid attention enables processing entire codebases, long documentation, or multi-document analysis without context truncation. GraphWalks evaluation shows meaningful reasoning performance even at 512k–1M tokens, critical for enterprise knowledge management.

Complex Software Engineering

35.7% on SWE-Bench (AgentLess), 75.6% on HumanEval+, strong MATH/GSM8K scores indicate capability for code generation, bug-fixing, and algorithm implementation. MTP 3x speedup aids rapid iteration in coding workflows.

Running & fine-tuning it

ESTIMATE—Verification recommended: FP8 mixed precision inference on 42B active parameters + KV-cache suggests ~80–150GB VRAM for dense batch inference (single batch, no quantization). Typical multi-GPU setup: 8× H100/A100 or 16–32× L40S/L40G. Single-GPU inference requires post-training quantization (INT8, GPTQ) reducing to ~40–60GB, with latency tradeoff. Training would require 64–256 GPUs depending on batch size. SGLang deployment example shows `--tp-size 16 --ep-size 16` for production serving, implying ~16-GPU minimum for reasonable throughput.

LoRA/QLoRA feasibility on MoE architecture is not explicitly documented. The 1.02T parameter count makes full fine-tuning impractical without industrial-scale infrastructure. Model card references domain-specific RL and MOPD multi-teacher distillation, suggesting post-training was complex, but no guidance on user-level adaptation is provided. Supervised fine-tuning via direct instruction-following appears possible given public SGLang deployment examples, but parameter-efficient tuning support is Unknown. Recommend contacting Xiaomi for LoRA compatibility and expert router stability during fine-tuning.

When to avoid it — and what to weigh

  • Latency-Critical Consumer Applications — MoE routing overhead and 1.02T total parameters require significant computational resources per inference. Multi-token prediction helps but does not eliminate latency penalty vs. smaller dense models. Not suitable for sub-100ms SLA requirements on standard hardware.
  • Resource-Constrained Environments — Even with 42B active parameters, sparse activation does not eliminate absolute VRAM and bandwidth requirements. Requires multi-GPU/multi-node deployment in most scenarios. Unsuitable for edge devices, mobile, or single-GPU inference without aggressive quantization.
  • Safety-Critical or Heavily Regulated Domains — No explicit safety audits, red-teaming results, or regulatory compliance details provided in model card. Post-training mentions 'safety teacher' but methodology and guarantees are not disclosed. Requires independent evaluation before deployment in healthcare, finance, legal, or autonomous systems.
  • Low-Resource Fine-Tuning — 1.02T parameter count makes full fine-tuning prohibitively expensive. LoRA/QLoRA feasibility on MoE architecture is unclear and not documented. Suitable only for organizations with substantial compute budgets for adaptation.

License & commercial use

MIT License (OSI-compliant, permissive). MIT allows unrestricted use, modification, and distribution, including commercial deployment, provided the license notice and copyright claim are retained. No additional restrictions stated on model card.

MIT license explicitly permits commercial use without royalties or additional licensing. No gating, API-only restrictions, or commercial partnership requirements stated. Model weights are publicly downloadable on HuggingFace and ModelScope. No affiliation or profit-sharing clause mentioned. Commercial use is clear from license standpoint. However, verify with Xiaomi if model output or derived models trigger any Xiaomi brand/trademark restrictions outside the license itself.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

No explicit security audit, red-teaming results, jailbreak resistance testing, or safety-tuning methodology disclosed. Model card mentions 'safety teacher' in MOPD training but does not detail safety constraints or guardrails. No disclosure of training data filtering for harmful content, bias audits, or RLHF safety alignment methodology. FP8 precision raises potential concerns for floating-point attacks, but no security analysis provided. Custom code trust required (model card shows `trust-remote-code` flag in deployment). Recommend independent red-teaming and responsible disclosure protocol before high-stakes deployment.

Alternatives to consider

DeepSeek-V4-Pro

Comparable 1.6T parameter MoE (49B active) with similar long-context and agentic focus. MMLU 90.1 (vs. MiMo 89.4), MATH 64.5 (vs. 86.2). Strengths: proven DeepSeek ecosystem; potential commercial backend. Weaknesses: unknown licensing model as of data cutoff, no published GraphWalks long-context evals.

Kimi-K2 (Moonshot AI)

1.04T parameter MoE (32B active) with 200k+ context claimed. MMLU-Pro 69.2 (vs. MiMo 68.5), HumanEval+ 84.8 (vs. MiMo 75.6). Strengths: proven long-context scaling in production; multilingual support. Weaknesses: availability/licensing unclear, limited base model eval disclosure, no open-source weights.

Llama 3.3 70B (Dense)

70B dense parameters, MIT license, well-documented, lower deployment friction. Benchmarks: MMLU 85.2 (vs. MiMo 89.4). Strengths: simple architecture, lower memory per token, stronger software ecosystem (vLLM, ollama.cpp). Weaknesses: short context (8k), no agent tuning, single-digit code benchmarks.

Software development agency

Ship MiMo-V2.5-Pro with senior software developers

MiMo-V2.5-Pro is ideal for teams building agent systems, long-context document analysis, or complex code workflows. Evaluate feasibility with our hardware calculator, review SGLang deployment best practices, and contact Xiaomi's community for production support. Start with a 16-GPU pilot on HuggingFace weights—no licensing friction.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

MiMo-V2.5-Pro FAQ

Can I use MiMo-V2.5-Pro commercially?
Yes. The MIT license explicitly permits commercial use, including in proprietary products and services, without royalties or additional licensing from Xiaomi. You must retain the MIT license notice in derivative works. No gating or API-only restrictions apply—weights are freely downloadable.
What is the recommended GPU setup for inference?
For production serving, SGLang documentation shows `--tp-size 16 --ep-size 16`, suggesting ~16 GPUs minimum. For research/testing with smaller batch sizes, 8× H100/A100 or 16× L40S is feasible. Single-GPU inference requires post-training quantization (INT8, GPTQ) and will be slow (~0.5–1 token/sec). Context length beyond 256k further increases VRAM.
Does MiMo-V2.5-Pro support fine-tuning or LoRA?
Full fine-tuning is impractical due to parameter count. LoRA/QLoRA feasibility on the MoE architecture is not documented in the model card. Supervised fine-tuning appears possible given public deployment examples, but parameter-efficient tuning support is unclear. Contact Xiaomi or the community for LoRA stability guidance.
How does the 1M token context actually work in practice?
The hybrid attention architecture (6:1 sliding-window to global) + learnable attention sink bias enables 1M token support without full quadratic complexity. GraphWalks evals show 0.37 BFS / 0.62 Parents accuracy at 1M, vs. ~0 for prior dense models. Real throughput and latency at 1M tokens are not disclosed; recommend benchmarking with SGLang on target hardware.

Custom software development services

DEV.co helps companies turn open-source tools like MiMo-V2.5-Pro into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Ready to Deploy MiMo-V2.5-Pro?

MiMo-V2.5-Pro is ideal for teams building agent systems, long-context document analysis, or complex code workflows. Evaluate feasibility with our hardware calculator, review SGLang deployment best practices, and contact Xiaomi's community for production support. Start with a 16-GPU pilot on HuggingFace weights—no licensing friction.