A Phone Runs 35 Billion Parameters, Moving From Demo to Daily Life

A Phone Runs 35 Billion Parameters, Moving From Demo to Daily Life

On October 13, ModelBest released a new generation of on-device multimodal large model, MiniCPM-V4.7, offering two versions at once: 35B-A3B and 1B. At first glance this looks like a routine iteration, but the 35B-A3B model deserves a closer look: 35 billion total parameters, built on a MoE (Mixture of Experts) architecture, activating only about 3 billion parameters per inference. The company calls it “the first open-source MoE model that runs losslessly and smoothly on the edge,” and it uses only about 3GB of memory when running.

Tested on an iPad Pro, decoding speed exceeds 15 token/s. You ask a question, it answers, with no need to watch a progress bar crawl.

The Dilemma of MoE on the Edge

MoE is popular because it does the math well: split the model into many “experts” and wake up only a few at a time. Total parameters can be huge while actual computation stays close to that of a small model. This path is comfortable in the cloud, where data centers have plenty of VRAM.

Moving to the edge gets awkward. The bottleneck is not compute, it is memory. Every expert has to be preloaded, because no one knows which one the next token will need. 35 billion parameters sitting there at normal precision means tens of GB, and phone RAM runs out fast. On the edge, “capable” and “able to run” become a single-choice question.

Waking Up the Next Expert in Advance

ModelBest's solution is called LES, Lookahead Expert Scheduling. The idea is not mysterious: if you can predict which experts will be used, then predict in advance. While the current matrix computation is still running, the system decides which experts will be invoked next and loads them into memory ahead of time, so loading and computing overlap.

Put another way, it changes “running to the warehouse for goods at the last minute” into “someone hands you the next item while you keep working.” The company says this is currently the industry's only open-source approach of its kind.

Two Versions, Each Covering Its Own Ground

35B-A3B is responsible for bringing flagship-level capability to the edge. It targets scenarios like these: you want comprehension close to cloud level, but you are not willing to send your data out. Offline document Q&A, instant recognition from a photo, and voice assistants in no-network environments are all within range. The 1B version pushes efficiency to the limit, aiming to let more devices run it, not just a few flagship phones.

On launch day, MiniCPM-V4.7 completed adaptation for chip platforms including Rockchip, NVIDIA, Intel, and Apple, supports iOS, Android, and HarmonyOS, and released edge adaptation code along with mainstream inference frameworks. The subtext is clear: this is not a lab demo, but something developers can pick up and ship.

The Path Choice Behind 60 Million Downloads

The MiniCPM series, maintained jointly by ModelBest and the OpenBMB community, has surpassed 60 million cumulative downloads, covering text, vision, and speech. That number is not small in the open-source world, showing that the edge small-model line meets stable, real demand, not just vendor talk.

In the same week, domestic edge and open-source camps made other moves. Xiaomi open-sourced the MiMo-V2.6 series; its flagship Pro version has 1.02 trillion total parameters, about 42 billion activated, a 1-million-token context, follows a “recursive self-improvement” reinforcement learning path, and unusually disclosed its RL training costs. The Institute of Automation of the Chinese Academy of Sciences, together with the Wuhan Institute for Artificial Intelligence, open-sourced Zidong Taichu ZDTaichu5.0-9B, releasing not only the weights but a full spatial multimodal data production pipeline, from data ingestion and cleaning to 3D annotation and synthesis verification.

These projects share one thing: what they give is no longer just a model, but a method for “turning your own data into model capabilities.” For teams looking to land results in their own industry, the latter is often more valuable.

What the Edge Actually Solves

Cloud models are smarter, and that will not change in the short term. The value of the edge is not in absolute capability but in a few very practical things: stable latency, usability without a network, and data that never leaves the device. For people building enterprise internal tools, medical or industrial inspection, and educational hardware, the last two often matter more than a few extra benchmark points.

Keep the gap in view, too. 35B-A3B activates 3 billion parameters per pass, so its capability roughly falls in the small-to-mid model range, still not the same as top closed-source models. Lossless edge running solves “able to run,” which is not the same as “runs best.” Treat it as an offline, private, always-ready assistant and your expectations will be more accurate.

Final Thoughts

Over the past year, the main battlefield of the large-model race has been parameters and leaderboards. But once the parameter war reaches a certain point, “can it fit into the device in a user's hand” starts to become a new dividing line. What is interesting about MiniCPM-V4.7 is not that it refreshed some score again, but that it pushed “running MoE on the edge” from a slogan to a reproducible engineering solution.

The imagination of the edge is not about copying the cloud, but about corners the cloud cannot reach: underground server rooms, mountains without signal, and ordinary jobs that do not want to upload work materials elsewhere. These scenarios will not trend, but they are the real market for edge models.