ModelBest puts MiniCPM in a $2,999 desktop edge AI node
ModelBest's Pinea Pi pairs local MiniCPM models with Intel or Nvidia compute and integrated sensors ahead of a planned Kickstarter campaign.
By RuntimeWire Staff · Published
Primary source: PR Newswire
Why it matters
Pinea Pi is ModelBest's attempt to move beyond supplying compact models and own the hardware, runtime and agent layer that connects local AI to physical devices.

ModelBest co-founder and CEO Li Dahai is taking the MiniCPM model family into hardware with Pinea Pi, a desktop edge AI node with early-bird prices of $1,999 for Lite and $2,999 for Pro. In an August 25 release, Pinea Pi said a Kickstarter campaign would run from late September through late October, with $30 reservations opening in late August.
Before the hardware bet, Li worked on search technology across several generations of China's consumer internet. A Zhihu filing says he worked on search at Google China, led search technology at mobile app distributor Wandoujia and served as Zhihu's CTO from 2018 until June 2024. He became ModelBest's CEO in June 2023, shifting his attention toward compact models designed to run close to the user.
That history explains the strategy underneath Pinea Pi. ModelBest has spent years arguing that smaller, denser models can move useful AI from remote data centers into phones, vehicles, robots and personal computers. Pinea Pi turns that research thesis into a product ModelBest can sell directly, with the sensors, runtime and models already assembled.
What the $1,999 and $2,999 boxes include
Pinea Pi has announced two configurations: a Lite model with up to 180 TOPS of Intel AI performance and a Pro model with up to 275 TOPS of Nvidia AI performance. Pinea Pi says both will run Ubuntu and include a MiniCPM-based edge AI software stack.
Pinea Pi describes the product through three labels. "Agent-native" means a user can change hardware behavior with natural-language instructions. "Multimodal-native" covers the integrated cameras, microphones, speakers and MiniCPM models. "Edge-native" means inference runs locally, including when the device is disconnected from the internet. These are Pinea Pi's descriptions of its product.
The manufacturer-listed specifications on the Pinea Pi Lite product page include an Intel Core Ultra X7 358H processor, 64 GB of LPDDR5x memory, a 1 TB NVMe drive, Ubuntu 24.04 and up to 180 TOPS of combined AI performance. The page also lists a four-microphone array, stereo speakers, an AI status light, four USB-A ports, two USB-C ports, HDMI, DisplayPort and dual Gigabit Ethernet. Farther down, the page repeats some Pro marketing copy, including references to 275 TOPS and six microphones; the Lite details cited here come from the Key Features block at the top.
The manufacturer-listed specifications on the Pinea Pi Pro product page include Nvidia's Jetson AGX Orin 64 GB module, which Pinea Pi rates at up to 275 TOPS. The page also lists 64 GB of LPDDR5 memory, a 1080p camera, a six-microphone circular array, a 1 TB NVMe drive and Ubuntu 22.04 with the MiniCPM software stack. Pinea Pi positions this configuration for robotics and embodied-AI workloads.
Pinea Pi's stated performance figures are higher than those of several lower-cost maker platforms, although the products are not directly comparable. Raspberry Pi's AI HAT+ 2 offers 40 TOPS of INT4 inference and 8 GB of dedicated memory for $130, before adding a Raspberry Pi 5 and peripherals. Nvidia's Jetson Orin Nano Super offers up to 67 INT8 TOPS for $249. Qualcomm and Arduino's VENTUNO Q offers up to 40 dense TOPS alongside a real-time microcontroller.
Pinea Pi is charging for a finished enclosure, integrated audio and vision hardware, local models and a software layer intended to remove much of the assembly work. The premium depends on whether that integration saves developers enough time to justify paying workstation prices.
Piny puts a consumer face on the developer box
ModelBest is also trying to stretch Pinea Pi beyond prototyping. The built-in assistant, Piny, is supposed to recognize its owner by face and voice, maintain editable personal memory and communicate through synchronized light and audio feedback. Pinea Pi says that memory can be audited and exported.
The Pinea Pi homepage describes a broader hardware hub that can connect to cameras, sensors, robots and smart devices through open interfaces. A user could describe a desired behavior by voice and have Pinea Pi coordinate those connected systems. If the runtime works as advertised, ModelBest would own a layer above the MiniCPM models: the protocol and agent software that decide how local AI interacts with physical hardware.
Privacy is central to the pitch. Pinea Pi says inference stays on the device, an illuminated ring communicates its operating status and a physical offline control keeps the system disconnected. The launch materials also promise no token fees or recurring cloud subscription. Those are Pinea Pi's own claims about the product.
Kickstarter will test the polish as well as the demand
Pinea Pi's homepage says the device is coming to Kickstarter, and the August 25 announcement describes a $30 reservation. The exact reservation opening date and Kickstarter launch date are not established beyond the company's late-August and late-September windows.
The announcements give no customer delivery date. For now, the performance ratings, privacy promises and software capabilities remain company-reported.
ModelBest has capital behind the attempt. In April 2024, China Daily reported that ModelBest raised several hundred million yuan in a round led by Primavera Venture Partners and Huawei's Hubble Investment, with Beijing's AI Industry Investment Fund and Zhihu also participating. ModelBest said the proceeds would fund recruitment, computing and data infrastructure, and the training and application of its models.
Pinea Pi gives that model work a direct route to customers and a possible reference design for other hardware manufacturers. The Kickstarter campaign will measure whether developers want ModelBest to assemble the full local AI stack for them, and whether Pinea Pi can turn its listed configuration into a reliable product.
For Li, the device makes ModelBest's efficiency thesis tangible. MiniCPM no longer has to win solely as a downloadable model. Inside Pinea Pi, it becomes the intelligence layer for a machine that can see, hear and control other devices. The $2,999 question is whether developers prefer buying that integration or building it themselves.