Xiaomi Registers Six LLMs: Multi-Model Strategy is Key for On-Device AI

The most valuable information in this article is that Xiaomi filing six large models for its new phone isn't simple "stuffing," but reveals its technical route shifting from "single-model dependency" to "multi-model routing" in edge AI deployment. This directly determines the practicality and cost of future smartphone AI experiences.

First, look at the filing list:

• Xiaomi Large Model (Self-developed)
• Xiaomi HyperOS Vision Large Model
• Xiaomi HyperOS Perception Large Model
• Xiaomi MiMo Large Model
• Zhipu AI Large Model
• ERNIE Bot Large Model
• (DeepSeek, Tongyi Qianwen, etc., mentioned in leaks may have been filed simultaneously but not fully listed; visible in the image)

This isn't simple "adaptation" or "pre-installation." The act of filing itself means Xiaomi needs to clarify to regulators how these models exist on the device—possibly as pre-set inference engines or cloud-local hybrid architectures. But more critically, Xiaomi filed both self-developed and third-party models simultaneously, pointing to a core question: Should edge large models follow an "All in One" or "Model Routing" path?

[!note]

Current industry mainstream practices: Qualcomm and MediaTek push general-purpose AI engines supporting multi-model loading; Apple insists on self-developed small models (like Apple Intelligence) and strictly controls third-party access. Xiaomi's filing action implies it chose the Qualcomm-style route—inclusive, but at the cost of model fragmentation and memory usage.

Short Term: Multi-model Filing is the "Lowest Cost Ecosystem Placeholder"

From a technical implementation perspective, deploying multiple large models on smartphones first requires resolving memory and compute conflicts. Taking current mainstream edge deployment solutions as examples, a 7B-level model quantized to 4-bit takes about 3.5GB, but phones usually only have 8-12GB of usable memory, making it impossible to keep multiple large models resident simultaneously. Xiaomi's solution is likely on-demand loading + model routing:

• At startup, only load lightweight models (e.g., MiMo, estimated parameter count 1-3B)
• Dynamically switch based on task type: visual tasks call the HyperOS Vision model, dialogue tasks call ERNIE Bot or DeepSeek
• Cloud fallback: Complex tasks (like code generation) use cloud APIs

In the short term, the benefit of this scheme is "accepting all comers"—users can call mature products like ERNIE Bot and Tongyi Qianwen on Xiaomi phones without Xiaomi needing to self-develop all capabilities. This can quickly boost the marketing hype for "AI Phones," similar to when phone manufacturers pre-installed multiple browsers back in the day.

But the technical hidden dangers are also obvious:

1. Model Switching Latency: Loading to inference typically takes 1-2 seconds, resulting in a worse experience than a single resident model

2. Privacy Fragmentation: Different models have different data processing strategies, making unified user control difficult

3. Version Fragmentation: Third-party model updates depend on vendors; Xiaomi needs to maintain adaptation versions for multiple models

Long Term: Self-developed Large Models are the "Moat," Third-party Models are "Transition Interfaces"

Spreading out Xiaomi's filing list reveals a clear hierarchy:

Tier	Model	Positioning	Deployment Method
Core Self-dev	Xiaomi LM, HyperOS Vision/Perception, MiMo	System-level AI capabilities	Pre-set + Resident
Strategic Partner	Zhipu AI, ERNIE Bot	General dialogue/knowledge	On-demand loading
Observational Filing	DeepSeek, Tongyi Qianwen	Tech validation/Eco compatibility	Optional download

Key judgment: Xiaomi cannot rely on third-party models long-term to define core experiences. Because the ultimate competitiveness of edge AI lies in "system-level intelligence"—such as smart scheduling, privacy computing, and cross-app collaboration—which must be completed by self-developed models. Third-party models can only serve as "app store" style add-on services.

In the long run, Xiaomi will gradually narrow the number of external models, keeping only 1-2 strong partners (likely Zhipu AI and ERNIE Bot), connecting others via lightweight API gateways. Meanwhile, self-developed models will iterate continuously, covering more scenarios:

# Pseudo-code logic for Xiaomi edge model routing
def route_task(task_type, user_preference):
    if task_type == "system_control":
        return load_model("mi_mo")  # Self-developed lightweight
    elif task_type == "image_understanding":
        return load_model("mi_vision")  # Self-developed vision
    elif task_type == "creative_writing":
        if user_preference == "zhihu_rate":
            return load_model("glm")  # Zhipu
        else:
            return load_model("wenxin")  # Wenxin
    else:
        return cloud_api("deepseek")  # Cloud fallback

Trend Prediction: By End of 2025, Smartphone Edge Large Models Will Shift from "Competing on Quantity" to "Competing on Throughput"

My judgment is: **Xiaomi's filing this time is essentially

Original link: https://www.ithome.com/0/981/378.htm