Breakthrough in On-Device Model Update Challenge: Airabi's Differential Algorithm Gains International Academic Recognition

Recently, the latest research achievement in the iterative upgrade of edge-side large models was announced: ResComp, a differential upgrade algorithm for edge-side large models developed by AIrabi Intelligent Technology, was successfully accepted by the Pacific Rim International Conference on Artificial Intelligence (PRICAI). This marks global academic recognition of the technology in lightweight deployment and efficient updating. This breakthrough not only provides key technical support for the continuous iteration of edge-side large models but also accelerates the large-scale iterative application of AI on terminal devices.

Edge-side large modelEnter the high-frequency iteration phaseThe bottleneck in upgrading urgently needs to be overcome.

As large models continue to expand their application scenarios, the shift from cloud to edge has become a clear trend. Especially in the automotive industry, edge-side large models are rapidly being implemented in core scenarios such as assisted driving and intelligent cockpits, thanks to their advantages of low latency, high privacy, and offline availability. They are being applied to a wide range of areas, including in-car interaction, operational optimization, and customized private models.

However, the issues of model iteration and upgrade have also emerged. The traditional OTA (Over-the-Air) upgrade method appears inadequate when dealing with model files that can be several gigabytes or even larger. Due to the complex parameter structure of large models and the extreme data compression (quantization), relying solely on overall compression results in very low transmission efficiency. Differential upgrades have become an optimal solution for achieving efficient upgrades given the limited resources on the end side.

The traditional differential algorithms demonstrate three major shortcomings in the practice of updating large models on edge devices. The paper points out that traditional incremental encoding algorithms, due to their failure to adapt to the special structure of LLM parameter data, face three significant challenges when generating model patches: low compression rates, slow update speeds, and high memory consumption. These issues severely restrict the experience upgrade and functional iteration of AI on edge devices.

Airlab Proposes ResComp, a Specialized Differential Algorithm for Edge Large Models, Demonstrating Outstanding Performance in Real-World Tests

Facing this industry bottleneck, a residual-based difference algorithm, ResComp, is proposed.

The ResComp algorithm reconstructs the update logic of edge-side LLMs through "structural alignment + residual optimization," achieving three core technological breakthroughs.

Firstly, break the underlying assumption of traditional algorithms' "discrete matching" by directly aligning the weight structures of the old and new models, accurately capturing the parameter change patterns, thereby reducing redundant data volume from the source.

Secondly, introduce the residual sequence calculation mechanism, combined with the bzip3 compressor to form a dual gain of "structural optimization + efficient compression."

Thirdly, by innovatively incorporating a Run-Length Encoding (RLE) enhancement mechanism, the patching speed during actual deployment is further increased by 30%.

Measured performance is excellent, with differential package size reduced to below 22% of the open-source algorithm.

To verify technical reliability, the team conducted multi-scenario tests on several mainstream open-weight LLMs and Stable Diffusion-based image generation models. The results show that whether it is text-understanding LLMs or multimodal models, the ResComp algorithm consistently achieves the comprehensive advantages of "smaller patch size, lower memory usage, and faster update speed."

Compared with certain open-source algorithms, Ailaby's ResComp technology, through an intelligent differential algorithm framework, can accurately identify the differences between large model versions, generate minimal differential packages, and significantly shorten upgrade times, bringing a new solution to the industry. (Open-source information is provided at the end)

We selected the most popular models on the market and compared the file sizes and differential compression effects in both the original floating-point format and the quantized format. By comparing a certain open-source differential algorithm with Airabit ResComp, we obtained the following results:

Difference package size: Taking the DeepSeek 1.5B 4-bit model as an example, the difference package generated by ResComp is only 90MB, while the open-source algorithm requires 400MB. The Airbi difference package size is only 5.7% of the original package, which is 22.5% of the open-source algorithm's difference package. For the 8B 4-bit model, the difference package size for ResComp is 0.27GB, while the open-source algorithm requires 1.35GB. The Airbi difference package size is 5.2% of the original package, which is 19.9% of the open-source algorithm's difference package.

The difference in time: For example, with DeepSeek's 1.5B 4-bit model, the upgrade time of ResComp is 50% of the open-source algorithm. For the 8B 2-bit model, the upgrade time is 51% of the open-source algorithm.

(1.5B uncensored, i.e., thirdeyeai/DeepSeek-R1-Distill-Qwen-1.5B-uncensored)

(8B Huihui, namely huihui-ai/DeepSeek-R1-Distill-Llama-8B-abliterated)

The technical breakthrough by Airabi not only possesses academic significance but also has a wide range of industrial application prospects. Currently, the algorithm has been successfully commercialized and integrated into Airabi's differential upgrade standardized platform. It is gradually being applied to the company's core business lines, such as intelligent terminal AI solutions and the vehicle large model upgrade platform.

It is particularly noteworthy that in the context of the Internet of Vehicles, differential upgrades of large models on the device side can not only significantly reduce data traffic costs but also enable seamless updates for users, avoiding interruptions to in-car services during the upgrade process. Even in situations with unstable network signals or no connectivity, reliable upgrades can still be accomplished through local caching and differential restoration.

With the continuous improvement of edge computing hardware performance and advancements in model optimization technologies, the breadth and depth of on-device large model applications will continue to expand. As a key infrastructure supporting the continuous evolution of models, the maturity of differential upgrade technology will directly impact the deployment pace of AI applications and user experience.

The introduction and implementation of the AIlaBi ResComp algorithm provide the industry with an academically endorsed and empirically validated on-device model upgrade solution, driving the evolution of edge AI applications toward "lightweight deployment and high-frequency iteration."

Abstract Information: