--- title: "Honor's MagicLM 3.0 puts a model on the phone that learns after you buy it" date: 2026-11-01 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-honor-magiclm3 language: en --- # Honor's MagicLM 3.0 puts a model on the phone that learns after you buy it > Honor released MagicLM 3.0 (魔法大模型3.0) and MagicOS 10 in October 2025, an on-device, self-evolving assistant called YOYO that splits work between phone and cloud. You buy a phone, and three months later its assistant suddenly handles three thousand tasks it couldn't on day one. Honor is arguing that the model on the device should keep growing with you — not freeze the moment it leaves the factory. It is a different bet from rivals who treat the assistant as a fixed feature shipped and forgotten. In October 2025, Honor launched the Magic8 series and MagicOS 10 in Beijing, built around MagicLM 3.0 (魔法大模型3.0) and an on-device, "self-evolving" assistant named YOYO. The core idea is that the phone, not a distant data center, becomes the place an ordinary user meets a foundation model (基础模型) — and that the model should keep improving from real use. ## Key takeaways - Honor released MagicLM 3.0 (魔法大模型3.0) and MagicOS 10 in October 2025, built around an on-device, self-evolving assistant called YOYO. - The model is an end-cloud matrix: small models run on the phone for privacy-sensitive tasks, larger models run in the cloud for heavy lifting. - YOYO covers 3,000+ everyday scenarios (shopping, food, travel, health, work) and earned a Level-3 agent certification from the China Academy of Information and Communications Technology (中国信通院). - The Magic8 series starts at about ¥4,499 (≈ US$627); the Pro starts at ¥5,699 (≈ US$794). ## What "self-evolving on the edge" means Honor frames MagicLM 3.0 as a model that "evolves" because the YOYO agent expands its reachable tasks over time. Honor says YOYO could handle only about 200 scenarios in July 2025 and roughly 3,000 by the Magic8 launch three months later — a fifteen-fold jump driven by adding agent capabilities rather than retraining the base model from scratch. The base model itself has progressed in clear steps: MagicLM 1.0 arrived in 2023 as an on-device, billion-parameter-class model on the Magic6; MagicLM 2.0 in 2024 introduced a four-layer, end-cloud architecture; MagicLM 3.0 adds multimodal understanding, a personal knowledge base, and autonomous execution. The throughline is moving intelligence onto the device while keeping the cloud for the heavy jobs. ## Why split the model across phone and cloud The split is the whole point. Sensitive, personal, and latency-critical tasks — remembering a preference, reading what's on screen, controlling the device — run on small models on the phone. Tasks needing broad knowledge or heavy reasoning go to larger cloud models. Honor says this keeps a user's private data on the device while still delivering capable answers. Honor's product chief put the privacy line plainly: the end side carries a user's private data and context, which is the user's core asset. That framing matters because an assistant that books, buys, and reads for you is also an assistant that knows a great deal about you. ## The agent that actually does things YOYO's headline ability is execution, not chat. Honor says it can complete tasks across more than 3,000 scenarios spanning "clothing, food, housing, travel, and shopping," and that it earned a Level-3 "excellent" agent certification from the China Academy of Information and Communications Technology (中国信通院), which Honor describes as the top current level for agent generalization. The phone includes a dedicated AI button that launches YOYO's vision mode — a live video call with the assistant — so it can act on what the camera sees: identify an item, find a product, or walk through a task. The pitch is an assistant that perceives, plans, calls tools, and executes, rather than one that only answers. According to Li An, chief scientist at BrainNet (脑机网), the edge-and-cloud split reflects a broader shift where the phone, not the data center, becomes the primary place ordinary users meet a foundation model. ## Beyond the phone MagicLM 3.0 is one piece of Honor's "Alpha Strategy," announced in early 2025 with the goal of becoming a global AI-terminal ecosystem company rather than a handset maker. The model matrix already spans language, speech, vision, and GUI-action models, some running on-device (such as a vision model Honor says uses low-bit quantization to cut storage and power) and some in the cloud for heavier tasks. Honor has also previewed a "ROBOTPHONE" — a device pairing an AI brain with robot mobility — planned for a 2026 reveal in Barcelona. That ambition matters for the model story: a foundation model that lives on a phone today is being positioned as the same intelligence that will drive other hardware tomorrow. The bet is that the assistant users train by daily use becomes portable across devices. ## Honest limitations The scenario-count growth (200 to 3,000) and the Level-3 certification come from Honor's launch materials and Chinese state-media reporting; we did not verify the certification independently. Marketing claims that Honor's on-device models beat specific foreign models were not reproduced and are omitted here. The Magic8 pricing is for the China launch and varies by market and configuration. On-device model quality depends heavily on the specific chip and memory in each phone, so experience will differ across the product line. We did not test YOYO's real task-success rate in daily use. ## What readers can do now - If you use a recent Honor phone, open MagicOS 10 and try YOYO on a real errand (book, compare, remember) to see whether on-device execution actually saves steps. - If you care about privacy, test which tasks stay on the device versus which call the cloud, and decide where your comfort line is. - If you follow edge AI, watch whether other phone makers match Honor's end-cloud agent pattern — it is becoming the default shape of on-device intelligence. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-honor-magiclm3 Free to quote with attribution and a link to the original.