What this is

Meta recently released MobileMoE, a family of “on-device MoE models.” On-device means the model runs directly on your phone or laptop without relying on the cloud. MoE (Mixture of Experts) is a network architecture that splits a large model into dozens of “small experts”; during each inference, only a few are activated, allowing the model to deliver the capabilities of a much larger model with less compute.

MobileMoE comes in three scales: small, medium, and large, with 0.3B, 0.5B, and 0.9B active parameters, corresponding to 1.3B, 2.8B, and 5.3B total parameters. Each tier is available in three variants: Base (the pretrained foundation model), SFT (supervised fine-tuning, or another training pass using human-labeled data), and QAT (quantization-aware training, which makes the model inherently suited to low-precision compression). The largest L version is less than 3GB after INT4 (4-bit integer) compression, so it can fit in the memory of an ordinary smartphone.

Industry view

Optimists see this as a “turning point for on-device AI.” In the past, getting capabilities close to those of a large model on a phone meant either going to the cloud (expensive, latency-prone, and a privacy concern) or using distillation (compressing a large model’s capabilities into a small model), which caused a drop in quality. MobileMoE combines MoE with fine-grained expert routing and shared experts, allowing a small number of active parameters to maintain quality.

We also note several important limitations:

  1. The license is FAIR NC (noncommercial): personal research is allowed, but products are not. Teams looking to integrate it into their own apps will have to sit on the sidelines.
  2. The context window is only 8,192 tokens (about 6,000 Chinese characters), which is not enough for long-document analysis.
  3. The training data is overwhelmingly English, so Chinese performance is unknown; Meta has not published multilingual benchmarks.
  4. Community-run real-world comparisons with closed-source small models of similar size from Qwen and GLM have not yet produced data.

Impact on regular people

For enterprise IT: Don’t expect to use MobileMoE as a product in the short term, but we recommend tracking this technical route. Over the next year or two, on-device AI will likely become cheaper and offer better privacy than cloud APIs, so projects such as customer service and document processing deserve another architecture review.

For individual professionals: The direct impact on day-to-day work is limited for now. But understanding the trend toward on-device AI is valuable in itself—the next wave of AI capabilities may not live in a web-based ChatGPT; it could run locally on your phone, available offline.

For the consumer market: The real beneficiaries are smartphone and operating-system vendors (Apple Intelligence and Google Gemini Nano). Meta has laid out a roadmap for “small but powerful,” putting pressure on the entire industry to accelerate.