As part of the Qualcomm Snapdragon Summit 2025, Honour has announced that the upcoming Magic8 family of smartphones will ship with the new fifth-generation Snapdragon 8 Elite chipsets. Thanks to the latter, Honor continues to define what AI in a smartphone will mean for its consumers..
Honor’s partnership with Qualcomm offers new interactions to consumers with benefits in computing power, storage, and battery consumption. Yet the focus feels geared towards a fourth goal that is a quiet desire of every manufacturer and consumer… efficiency. That’s not just about efficiency in hardware and software, but also about offering efficiency to consumers when they use their phones.
That comes with tradeoffs that Honor will be looking to balance out when it releases the features alongside the Magic8 family of smartphones.
The Benefits Of Compressing AI
The first appearance of large model AI features on smartphones required connections to the cloud. That remains the case for many features, but in 2025, there is a move to introduce as much processing on-device as possible. That’s more efficient in terms of data usage and energy requirements. This requires hardware and software to deliver solutions that meet those expectations.
AI models are large. Making those models smaller reduces the memory and storage required to operate the models, which is critical in any device, but especially in smartphones.
Honor is introducing low-bit quantisation technology into the Magic8 series. In broad strokes, this is a compression technology that reduces the size of an AI model, with Honor highlighting a 4.4 GB model dropping to 1.02 GB. That, in turn, allows calculations to be performed more efficiently, offering faster speeds and reduced power consumption, all while enhancing privacy by keeping the processing on the consumer’s smartphone.
It’s not without issues, though. While quantization offers significant efficiency gains, the trade-off is in a potential loss of accuracy. Compressing a model can impact the accuracy of the results. This can be compensated for during the model’s development, but it remains a significant challenge.
AI hallucinations are a well-known phenomenon to the public. These can quickly break down trust in an AI. Honor is balancing the risk of consumers losing confidence in their phone with a more efficient local AI model that can offer more results. I imagine that Honor will have guardrails on the use of the quantized model to reduce the chance of hallucinations by giving it tightly defined roles and functions in the OS.
Athina AI’s Himanshu Bamoria highlights that the loss of accuracy can be mitigated, and the lower relative performance of a smartphone-based model compared to the cloud may not have a notable impact:
“Surprisingly, for many common use cases, the loss was negligible. The quantized models maintained competitive accuracy across benchmarks like language understanding and reasoning tasks. This suggests that, for a vast range of applications, the extra computational burden of high-precision models might not be justified.“
The Power Of Two Local Searches
Searching across all of a smartphone’s content has been a standard feature for many years; why else carry personal information in a personal device if not to recall it at a moment’s notice?
Navigating information, rather than navigating apps, will be a key piece of infrastructure as smartphones transition from an app-first model to an AI-first model. New search tools are needed, and Honor is introducing one such option today: Next-Generation Hybrid Retrieval Technology.
The hybrid in this wordy title refers to the use of both standard keyword-based search tools and semantic search. Keyword-based search focuses on finding an exact match, whereas semantic search is more meaning-based, working effectively with inaccurate or vague search terms. Balancing these two approaches into a blended solution should yield more relevant results for consumers.
As part of a deep dive on hybrid search, Meilisearch’s Ilia Markov notes the drawbacks of this approach:
“Hybrid search requires integrating multiple search algorithms (e.g., semantic search using dense embeddings). This integration can be technically complex and requires deep technical understanding.”
Finding the balance point between these two approaches will be crucial. The results will, undoubtedly, be presented as a single, merged list. Both methods should be considered for inclusion without relying too heavily on one or the other. We’re returning to Honor looking to find the balance between trust and accuracy. For all the testing in a developer’s lab, when this hits the real world, consumers are going to have wildly different data sets from the lab, and they’re not going to want false results.
The Practical Applications Of Local Models And Improved Search
There is always a danger when presenting new techniques and approaches that the practicality is hidden behind the process. Honor is addressing this through Magic Color. With a single natural-language prompt, Magic Color should be able to search and pick out the desired photo, lift out key elements of color and tone, and re-style a second photo or video in the style of the first. No complicated tools, no scrolling through a photo-roll, and no working through many small tweaks in a digital photo studio. Just ask.
Honor has highlighted a series of prompts that cross over all of the new features; “Save the picture as a La La Land Color template” and “Help me find the photos I took at the Forbidden City” are combined into the final image.
Another area Honor has highlighted is in enhancing the visual quality of on-screen graphics. Primarily focused on gaming, it will balance the GPU and NPU to create an AI-powered Super Sampler. This should enable low-resolution gaming, running at lower frame rates, to be upscaled to HD resolution with frame rates of up to 120 FPS.
Various techniques are currently available that use the GPU for this upscaling, but Honor is using both the GPU and NPU in tandem as a “dual engine” to offer faster processing and lower power requirements.
Honor’s Alpha Plan At Work
Every smartphone manufacturer is in a race to bring not just AI into their smartphones, but to tilt the market towards their view of what AI should be. With the launch of the Magic8 series, Honor is making a statement that it is working on more than adding AI features.
Through the partnership with Qualcomm, it is reworking the core engineering of its smartphones. The efficiency of AI onboard the Magic smartphones and tablets is focused on bringing near-instantaneous personalized results to consumers while maintaining their privacy.
Now read more about Honor’s work to define AI through its Alpha Plan…
Read the full article here
