Anyone who travels frequently knows a simple truth: the capacity of a backpack is limited, but what you want to bring always exceeds its size.
Edge AI is facing the same dilemma. Over the past two years, user demand for edge intelligence has grown rapidly: editing photos right after taking them, generating instant summaries after recording videos, and extracting meeting minutes in real time. These scenarios all require local multimodal capabilities on devices. Smartphone manufacturers, chipmakers, and app developers are all stepping up their efforts, turning edge AI from an exclusive selling point of flagship models into a standard feature for smartphones.
However, hardware space is becoming increasingly tight. Data from Counterpoint shows that in Q2 2026, the quarter-over-quarter price increase of smartphone memory exceeded 80%. The proportion of memory chips in the bill of materials (BOM) cost of smartphones has climbed from 10% to 15% in the early years to 30% to 50%, with some mid-to-low-end models even reaching 60%.
On the other hand, IDC expects that in 2026, the shipment volume of next-generation AI smartphones in China will reach 147 million units, a year-over-year increase of 31.6%, with the penetration rate breaking through 53% for the first time. For every two new smartphones sold, one is equipped with edge AI. Globally, the penetration rate of generative AI smartphones will reach 45% in 2026 and exceed 52% in 2027, officially becoming an industry standard.
Demand is exploding, but the cost of memory—the 'backpack'—continues to rise. To achieve large-scale deployment of edge AI, the approach of simply stacking memory and parameters is no longer viable.
Against this backdrop, MiniCPM-V 4.7, recently released by Modelbest and OpenBMB, offers an ultra-minimalist backpacker-style solution—maximizing organization while keeping the existing backpack capacity. As the first open-source MoE (Mixture of Experts) multimodal large model that can run losslessly and smoothly on edge devices, this 35B-parameter MoE multimodal large model can run smoothly on systems like iOS and Android with only about 3GB of memory.
Looking closer, this is a layered, orderly, and meticulously calculated philosophy of organization.
If you have ever traveled, you probably understand this feeling:
Carrying a large suitcase is too heavy, a small backpack cannot hold everything, stuffing it full means spending half a day digging for things, and the zipper might burst open at any time.
Edge AI is currently facing the dilemma of these three types of bags.
The first type: The large suitcase in the cloud, powerful but impossible to carry.
Cloud large models are like a giant suitcase that can hold everything. GPT-4V-level multimodal capabilities excel at looking at images, reading documents, solving math problems, and analyzing videos. But the problem is that it requires an internet connection, has latency, raises privacy concerns, and incurs cost pressures. For terminal scenarios such as smartphones, tablets, and in-vehicle systems, what you often need is an AI that is readily available at your fingertips, not a large suitcase that needs to be checked every time.
The second type: Edge small models are like a waist bag, lightweight but unable to hold much.
So the industry has swung to the other extreme: making models smaller and stuffing them into smartphones. Hundreds of megabytes, 1B, and 2B edge models are emerging one after another. They can indeed run and are lightweight, but their capabilities are truly limited. OCR cannot accurately recognize complex tables, math problems are solved incorrectly, and video understanding only gets the general idea. It is like a waist bag; it is fine for holding a phone, keys, and a wallet, but if you really need to go out and get things done, the essentials are simply not enough. The user experience is that edge AI is usable, but not user-friendly.
The third type: A hiking backpack forced with excessive gear, able to hold a lot but too heavy to carry.
So, can we make edge models a bit larger? Of course, but the cost is memory.
Data from Counterpoint shows that in Q2 2026, DRAM memory has officially surpassed the main SoC (System on Chip) to become the most expensive single component in flagship smartphones. The price of 64GB DIMM has increased by about 3.5 times within a year, and the upward trend has not stopped. 12GB has become the passing grade for AI smartphones, 16GB is the standard for flagships, and some models are exploring 24GB. The more memory is stacked, the higher the cost, but consumers' wallets are not infinite. Relying on stacking memory to stack AI capabilities is becoming an increasingly expensive path. The cloud suitcase cannot be carried, the edge waist bag is not enough, and the gear-stuffed hiking backpack is too heavy to carry.
Edge AI seems to be stuck in an awkward middle ground.
Then, what kind of edge AI can meet current market and technical demands?
The 'Tau Scaling Law' proposed by Modelbest and Tsinghua University has long made a judgment: the capability density of large models doubles approximately every 3.5 months, and every 100 days or so, current optimal performance can be achieved with half the parameters. This means that efficiency is the true moat for edge AI.
And MiniCPM-V 4.7, recently released by Modelbest, provides a new perspective. The most striking feature of MiniCPM-V 4.7 is that it pushes the upper limit of multimodal models that can run smoothly on edge devices to a tier previously considered 'not for smartphones': a 35B total parameter, 3B activated MoE architecture, which can run smoothly on iOS and Android systems with about 3GB of memory.
This release simultaneously launches two versions: the 35B-A3B (35B total parameters / 3B activated) is the most powerful multimodal large model that can run smoothly on edge devices, capable of running smoothly on systems like iOS and Android with about 3GB of memory; the other 1B version (1.3B parameters) achieves first-tier performance in OCR, math, and video at the same size with an extremely small footprint. The entire series supports a 256K ultra-long context, and the model weights, inference code, and fine-tuning code are all open-sourced.
We can understand its art of organization from the following perspectives.
First layer of organization: 'Packing' the large model into small bags.
Experienced backpackers know a basic principle: do not stuff everything into one large bag, but use multiple packing bags to organize, taking out only what is needed. The MoE architecture applies this idea to large models. Traditional dense models are like a giant suitcase where all parameters must reside in memory simultaneously, and all must be utilized for every inference. In contrast, the MoE model splits the 35B total parameters into multiple 'expert modules,' activating only a small portion for each inference. This reduces computational volume and memory bandwidth requirements while maintaining the capacity and capabilities of the large model.
The activated parameters of MiniCPM-V 4.7 35B-A3B are only 3B. This is like having 35 pieces of gear, but only carrying the 3 most commonly used ones on your body when going out, with the rest packed in different storage bags to be taken out when needed. The total capacity is not reduced at all, but the carry-on burden is greatly lightened.
This is the first layer of guarantee for the high-efficiency route of MiniCPM-V 4.7: trading sparse activation for memory space. Memory usage drops significantly, yet the model's capabilities do not shrink. Efficiency and effectiveness are simultaneously achieved at this layer.
Second layer of organization: Anticipating in advance and putting what is needed within reach first.
Packing is just the first step. If you have to turn your backpack upside down every time you need something and then slowly dig through the packing bags, the efficiency would be too low. A true master of organization will anticipate what they need next and move it from the depths of the backpack to the outermost pocket in advance. The expert preloading technology of MiniCPM-V 4.7 does exactly this. Modelbest has proposed the Lookahead Expert Scheduling (LES) technology for the first time globally, which directly feeds the routing input features of the current layer into the routing module of the next layer, anticipates the required experts and loads them in advance, and significantly reduces loading wait times combined with optimized expert storage footprint and layout.
LES includes expert prefetching, compute scheduling, and weight fetching. In the MoE model, each Transformer layer needs to select experts before executing computations, and expert selection depends on the output of the previous layer. When expert weights are stored on an SSD, reading can only begin after the router selects the experts. This disk reading time is stuck before the computation of each layer, becoming a critical bottleneck for inference speed.
It is like looking for the next piece of gear; you must first figure out which one you need, then reach into the backpack to dig for it, and only use it once found. You have to wait at every step. The idea of preloading is that while the experts of the current layer are performing GPU computations, it predicts the experts to be activated in the next layer, and independent threads read their weights from the SSD in parallel into pre-allocated buffer slots in memory. In this way, the weight reading for the next layer overlaps with the unfinished computation of the current layer, eliminating idle waiting. Moreover, this prediction does not require training any additional networks; it simply reuses the router of the next layer, passing the input of the current layer's router through the next layer's router computation, taking the top 8 with the highest scores. Coupled with the optimization of the expert computation order, it allows the preloaded hit experts to be computed first, while simultaneously loading the few missed experts in parallel, further reducing waiting. Finally, LES also reduces loading costs further through fine-grained quantization and layout optimization, making the weights occupy less space and be placed in more continuous positions.
Actual test data shows that the coverage rate for truly activated experts exceeds 80%. Deployed and tested on an iPad Pro M4 16GB, the decoding speed is improved by about 5 times on average, and the absolute speed on the actual device can reach 17 to 19 tokens/s, making it the most powerful multimodal large model that can run smoothly on edge devices currently.
Just like an experienced backpacker who knows they will definitely need to drink water and replenish energy after climbing the mountain, so they move the water bottle and energy bars from the bottom of the backpack to the side pocket on the way down, so that when they really need them, they can just reach out and grab them.
Third layer of organization: Lightweighting and batch retrieval, making fast even faster.
There is another trick to organization: pack things needed for the same scenario together, take them out at once, without digging back and forth.
For inference where expert weights reside on the SSD, expert reading, host round-trips, and graph scheduling are fixed overheads that must be paid anew for every generated token. DSpark, however, uses a lightweight draft model to propose multiple candidate tokens at once, which are verified in parallel in a single forward pass by the target model, accepting the longest prefix that is consistent with its distribution. The fixed overhead is paid only once and shared by these tokens, significantly decreasing when amortized over each token.
Unlike general parallel speculation, the DSpark draft accompanying MiniCPM-V 4.7 adopts a semi-autoregressive structure. The parallel backbone solves the entire block at once, and the output end then uses a low-rank Markov module with a rank of 256 to apply a transition bias to the logits of the current position based on the previous draft token and pass it bit by bit into a chain. In this way, without increasing the number of forward passes, the dependency relationship of tokens within the block is restored, preventing the acceptance rate from decaying rapidly along the positions.
Besides batch retrieval, another trick is to reduce the weight of the gear itself. MiniCPM-V 4.7 retains the series' original x4/x16 hybrid visual compression. After 16x compression, MMBench still maintains 90.9 and OCRBench still maintains 90.4. Visual tokens are compressed to one-sixteenth of the original, with almost no drop in accuracy. When the things themselves become lighter, the backpack is naturally easier to pack.
While reducing weight, the capabilities have not shrunk. Coupled with sharded image encoding technology, MiniCPM-V 4.7 continues the series' capability accumulation in high-frequency edge scenarios such as long document OCR extraction, continuous dynamic video logical reasoning, and multi-image context joint reasoning. These scenarios were often considered too difficult for small models in the past, yet 4.7 still makes them usable at an extremely small size.
In the past, you could only take out one thing from the backpack at a time; each retrieval meant opening the bag, searching, and closing it. Now you can take out a whole set of toiletries at once, taking out the toothbrush, toothpaste, and towel together. Tested on an M5 16GB Mac, in tasks such as code debugging and implementation, syntax correction, scientific explanation, logical reasoning, and mathematical calculation, the decoding speed TPS under the optimal n_max is improved by an average of 27% and up to 66%. Speed goes up and accuracy does not drop; this is the true 'making fast even faster' that edge devices really need.
Fourth layer of organization: Careful selection, knowing what to bring and what not to bring.
True masters of organization understand a principle: it is not good just to pack everything in. Sometimes, knowing what not to bring is more important than bringing everything.
The MiniCPM-V 4.7 1B version has only 1.3B parameters, yet it achieves leapfrog performance among models of the same size. Compared to Qwen3.5-0.8B, and even the larger LFM2.5-VL 1.6B/3B, it comprehensively outperforms them in MMStar, MathVista, OCR, chart reasoning, hallucination suppression, and video understanding, scoring 74.3 on MathVista and 86.1 on OCRBench, with video understanding far exceeding competing products in the same tier.
It is like an experienced backpacker who carefully examines every item before going out, and what is finally kept are the truly high-frequency and truly core things. OCR, math, video... these are the most commonly used capabilities for edge multimodal, and the 1B version lacks none of them, achieving the first tier at the same size. But those rarely used, space-consuming features are decisively left at home.
The four layers of organization progress step by step, pushing the most core capabilities to the extreme with an extremely small size. The backpack has not become larger, but it holds more things.
Achieving this level of organization, when placed in the industry context, has actually crossed a line that many manufacturers have not yet crossed.
Smartphone manufacturers have not failed to think about running large models on edge devices. Leading smartphone manufacturers such as vivo, Honor, OPPO, and Xiaomi are all betting on the direction of edge MoE. But as of now, the mainstream edge models that can run on smartphones in the industry are still stuck in the 3B to 7B dense model tier. Running 30B-level MoE on smartphones is on the timeline of two to three years later for most manufacturers.
MiniCPM-V 4.7 has advanced this timeline. With about 3GB of memory and an actual device speed of 17 to 19 tokens/s, it means that a 35B total parameter multimodal model can now run smoothly on mid-range models. Multimodal understanding capabilities such as complex document OCR, continuous video logical reasoning, and multi-image context joint reasoning, which were previously only available on flagship phones, have the technical conditions to be brought down to the mid-range price segment for the first time.
However, no matter how well the organization is done, it only packs things into one's own backpack. For edge AI to truly move towards large-scale deployment, another thing is needed: teaching the packing method to more people, so that every device and every manufacturer can pack their own share accordingly.
A good organization method, if known only to oneself, has limited value. If shared so that more people can use it, it can truly change an industry.
From the previous generation to this one, the iteration of the MiniCPM-V series itself is a process of continuously upgrading this organization philosophy. MiniCPM-V 4.6 achieved the best performance at the same size with 1.3B parameters and 6GB of memory; V4.7 directly pushes the upper limit of smoothly runnable models to the 35B MoE tier, while memory is instead reduced to 3GB. The parameter count has increased by dozens of times, yet the memory requirement is halved.
Why is this important? Because 3B to 7B models represent different capability boundaries. Edge devices equipped with 7B models can recognize simple instructions but cannot read complex documents; they can answer basic questions but cannot understand video logic. The industry calls this 'usable but not user-friendly.'
Over the past few years, the 'capability equity' of edge AI has mostly stayed at the slogan level. Models that flagship phones can run, mid-range phones cannot; models that mid-range phones can run lack sufficient capabilities. Memory cost is the wall standing in the middle. When 3GB of memory can carry a 35B MoE, a breach is chiseled into this wall. For budget smartphones and mid-range phones, there is the possibility of using flagship-level multimodal capabilities for the first time, and technology for all is no longer just a vision.
For smartphone manufacturers, chipmakers, and app developers, this means new product space. They no longer need to compromise model capabilities for memory budgets, nor do they need to reserve the best edge AI only for the most expensive models.
Moreover, MiniCPM-V4.7 open-sources the model weights, inference code, and fine-tuning code across the entire series. This means that developers and manufacturers can not only use them but also conduct secondary development and optimization based on their own scenarios.
To further facilitate secondary development, MiniCPM-V 4.7 simultaneously introduces Self-Replay Fine-Tuning technology. The approach is very intuitive: before fine-tuning, the model is first asked to answer a batch of questions covering its original capabilities, high-quality answers are filtered out, and then these answers are mixed into the training data for the new task. While learning new skills, the model reviews what it originally knew, reducing the forgetting of existing capabilities.
These 'review materials' come from the model itself, making them closer to its original answering style. The entire method follows the standard fine-tuning process without modifying the model structure or loss function, and the pre-generated replay data can also be reused in subsequent experiments. Developers can therefore expand new capabilities at a lower integration cost while retaining the model's original general performance as much as possible.
Behind the complete set of edge multimodal AI solutions of MiniCPM-V 4.7 is the continuous investment of the team led by Yuan Yao, Co-Chief Scientist of Modelbest. As an Assistant Professor and Doctoral Supervisor at the School of Artificial Intelligence, Tsinghua University, Yuan Yao has published more than 30 papers in top conferences and journals such as Nature Communications, CVPR, ACL, and ICML, and led the construction of efficient multimodal models such as MiniCPM-V and MiniCPM-o. From the very beginning to the recent 4.6 and 4.7, the MiniCPM-V series has always iterated around the main line of 'running stronger models with fewer resources.' This backpacker-style organization philosophy is the continuation of the team's consistent efficiency route in edge multimodal.
Modelbest has also established an extensive chip cooperation ecosystem. On the day of the 4.7 release, it was adapted to chip platforms such as Intel, Rockchip, AMD, ARM, Houmo Intelligent, NVIDIA, and Apple. Previously, it has established cooperation with dozens of chip manufacturers such as Qualcomm, MediaTek, Huawei Ascend, and UNISOC.
The MiniCPM series has been successively deployed for B-end customers in automobiles, PCs, smartphones, smart homes, etc., including Lenovo, Geely, SAIC Volkswagen, GAC, Mazda, and Hongqi. As of October 2026, the cumulative downloads of the MiniCPM series have exceeded 60 million times, topping the GitHub Trending and HuggingFace trending lists for multiple consecutive days, becoming one of the most highly attended multimodal edge open-source foundations globally.
Previously, related achievements of MiniCPM-V have been published in Nature Communications, proving that Modelbest's edge multimodal route has been recognized by the international top academic community. The industry influence of this route was evident even earlier.
In 2024, a team from Stanford University released a model named Llama3-V, claiming to achieve GPT-4V-level performance at an extremely low cost. However, developers soon found that its architecture and code were highly consistent with MiniCPM-Llama3-V 2.5, previously open-sourced by the MiniCPM team. Ultimately, Stanford team members publicly apologized and withdrew the model. The 'repackaging' incident shows that China's multimodal large model teams have gradually shaken off the identity of followers and begun to become a technological source directly referenced by global peers. The open-source work of Modelbest has entered the 'source supply side' of global open-source innovation.
The backpacker's organization philosophy is essentially maximizing efficiency through orderly organization and anticipation under the premise of limited resources. What MiniCPM-V 4.7 does is the same.
Using the MoE architecture to pack the large model into small bags, using expert preloading to anticipate the modules to be used in advance, using DSpark to process multiple tokens at once, and using hybrid visual compression to 'lightweight' visual information... layer upon layer of organization techniques are stacked together, ultimately achieving the result of running a 35B MoE multimodal large model in 3GB of memory.
AI is penetrating everyone's daily life at a speed far exceeding expectations, and the edge is the stretch of road where it can most reach ordinary people. In the past, the ship of edge AI has always been parked in the deep-water port of flagship phones; being able to run was considered good enough. Now, users demand that it runs fast, performs well, and occupies little memory. What MiniCPM-V 4.7 does is untie the mooring ropes, allowing it to sail into the waters of every mid-range device.
When a model can deliver 35B capabilities in 3GB of memory, the beneficiaries are not just flagship phone users, but also a large number of mid-range devices, overseas markets, and scenarios with rigid demands for privacy and offline capabilities. It can be said that Modelbest has delivered MiniCPM-V 4.7 to these devices through a backpacker-style minimalist journey. And the destination of this journey is the day when edge multimodal truly moves from being usable to being user-friendly.