Home/Models/Xiaomi/mimo-v2-pro
X

mimo-v2-pro

輸入:$0.8/M
輸出:$2.4/M
MiMo-V2-Pro 是 Xiaomi 的旗艦級基礎模型,具備超過 1T 的總參數與 1M 的上下文長度,並針對 Agent 化場景進行了深度優化。它可高度適配 OpenClaw 等通用 Agent 框架。在標準的 PinchBench 與 ClawBench 基準測試中名列全球頂尖,其感知表現接近 Opus 4.6。MiMo-V2-Pro 旨在作為 Agent 系統的大腦,編排複雜工作流程、推動生產級工程任務,並可靠地交付結果。
新
商業用途
Playground
概覽
功能
定價
API

Xiaomi MiMo-V2-Pro 的技術規格

ItemXiaomi MiMo-V2-Pro
ProviderXiaomi
Model IDmimo-v2-pro
Model familyMiMo-V2
Model type代理型基礎模型 / 推理模型
Primary input文本
Primary output文本
Context window最多 1,000,000 個 token
Total parameters超過 1 兆
Active parameters420 億
Architecture混合注意力 MoE
Release window2026 年 3 月
Benchmark signalArtificial Analysis Intelligence Index:全球 #8;PinchBench:全球 #3

什麼是 Xiaomi MiMo-V2-Pro?

Xiaomi MiMo-V2-Pro 是 Xiaomi 面向真實場景代理工作的旗艦 MiMo 模型。Xiaomi 將其描述為支撐代理系統的核心模型,能協調複雜工作流、處理生產級工程任務,並在冗長、多步驟的作業中保持可靠運行。

Xiaomi MiMo-V2-Pro 的主要特性

  • 以代理為先的設計: 相較僅提供聊天式回答,更側重於工作流、工具使用與任務執行。
  • 超長上下文: 支援最多 100 萬個 token,適用於龐大程式碼庫、長篇文件與長期任務軌跡。
  • 大規模 MoE: 總參數超過 1T,有效參數 42B,並結合混合注意力以提升效率。
  • 強大的編碼能力: Xiaomi 表示,其編碼表現於內部評估中超越 Claude 4.6 Sonnet。
  • 可靠的工具呼叫: Xiaomi 強調在代理框架中的工具呼叫穩定性與準確性有所提升。
  • 框架友好: Xiaomi 稱該模型正與 OpenClaw、OpenCode、KiloCode、Blackbox、Cline 等代理框架配套使用。

Xiaomi MiMo-V2-Pro 的基準表現

根據 Xiaomi 於 2026 年 3 月的資料,MiMo-V2-Pro 在 Artificial Analysis Intelligence Index 上位列全球 #8,在 PinchBench 平均任務完成率上位列全球 #3。Xiaomi 亦報告其 ClawEval 分數為 61.5,稱其接近 Claude Opus 4.6,並在該基準上領先 GPT-5.2。

Xiaomi MiMo-V2-Pro vs MiMo-V2-Flash vs MiMo-V2-Omni

ModelBest forKey difference
MiMo-V2-Flash快速、高效的文本推理更小的 MoE 模型,針對效率調校;總參數 309B / 有效參數 15B
MiMo-V2-Pro深度代理式推理與長流程旗艦文本代理模型,具備 1M-token 上下文與 1T+ 參數
MiMo-V2-Omni多模態理解與執行統一文本、視覺與語音以處理多模態代理任務

何時使用 Xiaomi MiMo-V2-Pro

當您需要長上下文推理、多步驟代理編排、偏重程式碼的工作流,或接近生產環境的任務執行時,請選擇 MiMo-V2-Pro。當深度比速度更重要時,它比 MiMo-V2-Flash 更適合;當工作負載以文本為先而非多模態時,它比 MiMo-V2-Omni 更合適。

限制

MiMo-V2-Pro 定位為以文本為先的代理模型,因此原生多模態工作更適合由 MiMo-V2-Omni 處理。與任何以基準為導向的模型一樣,實際結果仍取決於提示設計、工具品質,以及代理如何接入您的技術棧。

常見問題

What makes Xiaomi MiMo-V2-Pro API different from MiMo-V2-Flash?

MiMo-V2-Pro is Xiaomi’s flagship agentic model for deeper workflows, while MiMo-V2-Flash is the efficiency-focused sibling. Xiaomi says Pro is built for real-world agent tasks, with over 1 trillion total parameters, 42 billion active parameters, and a 1 million-token context window.

How large is the Xiaomi MiMo-V2-Pro API context window?

Xiaomi says MiMo-V2-Pro supports up to 1 million tokens of context. That is the key spec to know if you need to keep huge codebases, long documents, or extended task histories in one run.

Can Xiaomi MiMo-V2-Pro API handle coding and multi-step agent workflows?

Yes. Xiaomi positions MiMo-V2-Pro as a model for production engineering tasks, complex workflows, and agent scaffolds. The company also says its coding ability surpasses Claude 4.6 Sonnet in internal evaluations.

When should I use Xiaomi MiMo-V2-Pro API instead of MiMo-V2-Omni?

Use MiMo-V2-Pro when your workload is text-first and centered on reasoning, code, or tool orchestration. Use MiMo-V2-Omni when you need native multimodal understanding across text, vision, and speech.

How does Xiaomi MiMo-V2-Pro API compare with Claude Opus 4.6 and GPT-5.2?

Xiaomi reports MiMo-V2-Pro at 61.5 on ClawEval, compared with 66.3 for Claude Opus 4.6 and 50.0 for GPT-5.2 on the same chart. Xiaomi also says Pro is close to Opus 4.6 on general agent performance and ranks #8 globally on the Artificial Analysis Intelligence Index.

What are the known limitations of Xiaomi MiMo-V2-Pro API?

MiMo-V2-Pro is optimized for agentic text workflows, so it is not the family member to choose for native multimodal input. For image, video, or speech-heavy jobs, Xiaomi’s MiMo-V2-Omni is the better match.

How do I integrate Xiaomi MiMo-V2-Pro API with an OpenAI-compatible client?

OpenClaw documents the Xiaomi provider as OpenAI-compatible, which means you can use an OpenAI-style client with Xiaomi’s base URL and model ID. In practice, that makes it straightforward to swap in mimo-v2-pro as the model name while keeping your existing chat-completions flow.

Is Xiaomi MiMo-V2-Pro API suitable for long document analysis?

Yes. The 1 million-token context window makes MiMo-V2-Pro a strong fit for very long source documents, support tickets, policy packs, or repository-scale analysis where smaller-context models would truncate too early.

mimo-v2-pro 的功能

探索 mimo-v2-pro 的核心功能,專為提升效能和可用性而設計。了解這些功能如何為您的專案帶來效益並改善使用者體驗。

mimo-v2-pro 的定價

探索 mimo-v2-pro 的競爭性定價,專為滿足各種預算和使用需求而設計。我們靈活的方案確保您只需為實際使用量付費,讓您能夠隨著需求增長輕鬆擴展。了解 mimo-v2-pro 如何在保持成本可控的同時提升您的專案效果。
彗星價格 (USD / M Tokens)官方價格 (USD / M Tokens)折扣
輸入:$0.8/M
輸出:$2.4/M
輸入:$1/M
輸出:$3/M
-20%

mimo-v2-pro 的範例程式碼和 API

存取完整的範例程式碼和 API 資源,以簡化您的 mimo-v2-pro 整合流程。我們詳盡的文件提供逐步指引,協助您在專案中充分發揮 mimo-v2-pro 的潛力。
Python
JavaScript
Curl
from openai import OpenAI
import os

# Get your CometAPI key from https://api.cometapi.com/console/token, and paste it here
COMETAPI_KEY = os.environ.get("COMETAPI_KEY") or "<YOUR_COMETAPI_KEY>"

client = OpenAI(api_key=COMETAPI_KEY, base_url="https://api.cometapi.com/v1")

stream = client.chat.completions.create(
    model="mimo-v2-pro",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain the Monty Hall problem step by step."},
    ],
    stream=True,
    extra_body={"thinking": {"type": "enabled"}},
)

thinking = False
for chunk in stream:
    delta = chunk.choices[0].delta
    reasoning = (delta.model_extra or {}).get("reasoning_content")
    if reasoning:
        if not thinking:
            print("<thinking>")
            thinking = True
        print(reasoning, end="", flush=True)
    elif thinking and delta.content:
        print("
</thinking>
")
        thinking = False
        print(delta.content, end="", flush=True)
    elif delta.content:
        print(delta.content, end="", flush=True)

更多模型