Most "AI phone" marketing means a button that sends your request to a distant data center. vivo is pushing a different default: the model lives on the device, your data never leaves it, and the cloud is only called in for the heavy lifting.
What BlueLM actually is
vivo first released its self-developed BlueLM (蓝心大模型) on November 1, 2023, alongside OriginOS 4 and the BlueOS (蓝河) operating system. The broader "BlueHeart Intelligence" (蓝心智能) strategy — fusing large models with the phone OS — was laid out at the 2024 vivo Developer Conference. By 2025 the company had built a tiered agent model matrix: a real-time voice model (BlueLM-RealTime), an on-device model (BlueLM-Nano) for perception and personal memory, and cloud models for cross-app, cross-device tasks.
The centerpiece is BlueLM 3.5 Nano 3B, a 3-billion-parameter on-device model released at the October 10, 2025 vivo Developer Conference. Numbers that matter for a phone: memory footprint cut from 3.5 GB to 2 GB, output speed lifted to 200 tokens per second, a 128K context window, and built-in UI-agent capability. On the SuperCLUE OnDevice benchmark for phone-side models, it scored 89.86 in 2026 — first place in its parameter class.
The 30B MoE that may change the equation
The more interesting bet is in the lab. vivo is pre-researching a BlueLM 30B MoE (Mixture-of-Experts) on-device model: 30B total parameters, but only about 1B–3B active per task, running in just 2–4 GB of memory. Because only a sliver of the model activates, a phone can handle long-chain complex tasks with the cloud never touching the data. That is the "impossible triangle" — small model, strong capability, low power — that on-device AI has been chasing.
The system play: Harness and agents
vivo's differentiation is less a single model than the plumbing. A system-level BlueHeart Harness (蓝心Harness) sits inside the OS kernel and turns more than 6,000 atomic skills into tools an agent can call, supporting over 10,000 task types. At the 2026 vivo Developer Conference (September 16, Shenzhen), the company launched a refreshed BlueHeart Intelligence strategy, OriginOS 7, and BlueOS 4, and introduced an upgraded personal AI assistant, BlueLM Small V (蓝心小V) Pro, that can break a task into steps, pull information across apps, and operate across devices — one sentence can produce a travel plan or a cross-device report.
The safety framing is notable. With on-device models, generated content never passes a server, which historically meant no cloud moderation. China's regulator addressed this directly: on July 8, 2026, seven brands — Apple, Huawei, OPPO, vivo, Xiaomi, Samsung, and ZTE (including Nubia) — filed on-device model records (备案) on the same day, establishing a compliance framework for "AI living in the phone."
Why it matters
vivo reports nearly 600 million global users, so even small per-device AI improvements reach a massive base. The strategic point is ownership of the interaction layer. If the assistant that "hears you, remembers you, gets things done" (听懂你、记住你、办成事) runs locally and is wired into the OS, the phone maker — not a cloud model vendor — controls the relationship. That is why Apple, Samsung, and the Chinese OEMs are all racing here simultaneously.
There is also a privacy argument that actually holds up: a 30B MoE that activates 1B–3B on-device means chats, photos, and documents can be processed without leaving the handset. In an era of constant cloud-data anxiety, "your data never leaves the phone" is a real product difference, not a slogan.
What readers can do now
- If you own a recent vivo phone: open BlueLM Small V and test the cross-app tasks — travel planning, report整理 — to see how far on-device agents have come versus cloud-only assistants.
- If you care about privacy: ask any "AI phone" vendor whether inference runs locally or relays to a cloud. vivo's 30B-MoE-in-2GB plan is the honest version of "on-device"; many rivals still relay.
- If you build mobile AI: watch the SuperCLUE OnDevice leaderboard. The 89.86 score is the bar small on-device models are being measured against in 2026.
Honest limitations
Model sizes, memory footprints, token speeds, and benchmark scores (89.86 on SuperCLUE OnDevice) are vivo's disclosures at its 2025 and 2026 Developer Conferences, reported through state and industry media (Economic Information Daily, Economic Daily, China Economic News); they were not independently benchmarked here. The 30B MoE model is described as "pre-research" (预研), meaning it is not a shipping product and timelines are uncertain. The "nearly 600 million global users" figure is vivo's own. The July 8, 2026 on-device-model filing by seven brands is reported industry news; the compliance framework's practical effect is still emerging. Performance claims should be read as vendor-reported, not third-party audited.