On-device AI

Underdog

2026-10-03 14:49:18

Sigil Wen launches Underdog, a private personal AI built to run on users’ own devices

Conway Research founder Sigil Wen has announced the launch of Underdog, described as a private personal AI that runs on devices users already own. According to the announcement, the project has backing from a16z, Khosla Ventures, Hummingbird VC, and Anthology, a vehicle tied to Anthropic and Menlo Ventures, along with support from several figures in the AI field including Patrick Collison, Naval, Guillermo Rauch, Noam Brown, Thomas Wolf, and Logan Kilpatrick. Underdog says its mission is to deliver free, usable, reliable, and private AI to billions of people. The team also put forward what it calls “Underdog’s Law,” arguing that frontier intelligence available today will reach user devices within six months. The project plans to begin with high-speed AI that can run on consumer hardware and aims to improve capability, speed, power use, and data efficiency through joint design of models and inference engines. Its research scope includes agentic commerce, confidential inference, and how AI could reshape the internet economy.

70

Tether

2026-09-22 00:06:30

Paolo Ardoino to host Reddit AMA on Oct. 7 focused on on-device AI and decentralized technology

Tether Chief Technology Officer Paolo Ardoino said on X that he will hold a Reddit AMA on Oct. 7 at 3:00 p.m. UTC. The session is set to cover on-device AI runtimes, privacy, local inference, and the future of decentralized technology. In the same post, Ardoino said Tether had previously addressed financial inclusion through USDT and is now tackling what he described as intelligence inclusion through the QVAC project. He also invited community members to join the discussion. The announcement was attributed to Ardoino’s post on X, cited by Techub News.

260

Cactus Comput

2026-09-19 04:57:44

Cactus Compute unveils Needle 3, a compact on-device model built for tool calling

Cactus Compute, a Y Combinator S25 on-device AI startup, has introduced Needle 3, a model designed for tool calling and structured extraction rather than general chat. In the example cited by BlockBeats, when a user says "set the living room light to 30%," the model identifies the right function and fills in parameters such as "living room" and "30%." The report said Needle 3 resembles the recently popular Jev in some ways, but follows a different path: Jev scores candidate decisions directly, while Needle 3 still generates token by token, with outputs constrained to tool calls that software can execute directly. One set of Needle 3 weights can be trimmed from 2 to 20 layers. The 2-layer version is about 9 MB, while the full 20-layer model is about 35 MB, allowing deployment choices across devices including watches, Raspberry Pi boards, and smartphones. After separate fine-tuning for Android operation commands, the 4-layer, 29 million-parameter version posted a self-tested score of 62.5, above DeepSeek V4 Flash at 60.5.

430

MiniCPM5-2B

2026-09-08 05:13:22

MiniCPM5-2B Open-Sourced: 2B Model Tops AI Benchmark Under 4B Parameters

Mianbi AI (Facewall Intelligence) and the OpenBMB community have officially open-sourced the MiniCPM5-2B, a 2-billion-parameter text model designed for local devices such as smartphones and PCs. The model natively supports a context length of approximately 128K tokens. First previewed in July, the release now includes the full weights, training data, training recipe, and reinforcement learning (RL) system, all under the Apache 2.0 license. In independent benchmarks by Artificial Analysis, the MiniCPM5-2B scored 15 on the Intelligence Index v4.2, ranking first among all open-weight models with fewer than 4B parameters, ahead of the second-place Granite 4.2 3B (11 points). This open-source release provides developers with comprehensive resources to deploy and customize the model for on-device AI applications, demonstrating the potential of small-scale models in achieving competitive performance.

810

Samsung Elect

2026-09-07 02:09:06

Samsung and Arm said to be developing a 2 nm on-device AI chip, with OpenAI rumored as a possible customer

Samsung Electronics and Arm are reportedly working together on a next-generation AI chip built on Samsung’s 2 nm gate-all-around (GAA) process, according to reports cited by ABMedia from outlets including TradingKey. The project is focused on a system-on-chip for on-device AI, and Arm reportedly approved the non-recurring engineering, or NRE, cost for the program in late August, marking the formal start of joint development. The reported division of work gives Arm responsibility for the AI accelerator architecture and register-transfer level design. Samsung’s System LSI unit is said to be handling SoC design and integration using Arm’s IP, while Samsung Foundry would take charge of manufacturing, advanced packaging, and commercialization. For Samsung Foundry, the project is framed as an important step in winning a flagship customer at the 2 nm node. One of the main questions around the project is the end customer. Industry sources cited in the report said OpenAI is a candidate, but Arm, Samsung, and OpenAI have not confirmed that claim. ABMedia also noted that Samsung has recently been active in custom AI chip efforts, and that earlier reporting had linked Anthropic with discussions around a custom 2 nm AI chip with Samsung.

840

Arm

2026-09-04 19:38:55

Arm and Samsung Team Up on 2nm AI Chip Project

Arm and Samsung Electronics have announced a collaboration to develop AI chips based on a 2nm process. The partnership aims to enhance on-device AI capabilities, reduce reliance on cloud services, and improve data privacy.

780

Samsung Elect

2026-09-04 10:09:58

Samsung and Arm Launch Joint Development of On-Device AI Chip

According to a report from South Korean media Greened.kr, as cited by ChainCatcher, Samsung Electronics and Arm have officially initiated a joint project to develop an on-device AI SoC chip. Arm approved the non-recurring engineering (NRE) fee at the end of August, and the project has since commenced. The collaboration follows a model where Arm provides core technology and Samsung Electronics integrates and delivers the final customized chips to end customers. The partnership focuses on developing customized SoCs for terminal devices, leveraging Arm's core IP and Samsung's integration capabilities. This project aims to produce tailored chips for edge-side AI applications, with the goal of delivering customized solutions to end customers.

860

Violoop

2026-09-01 13:16:19

Violoop raises an eight-figure yuan angel and pre-A round to build a hardware-based personal AI agent

Shenzhen-based AI hardware startup Violoop has completed an angel round and a pre-A round worth more than RMB 100 million, according to IT Juzi, with backing from Lenovo Capital, CICC Porsche, BlueRun Ventures, Genesis Capital, Qifu Capital and Zero2IPO Holdings. Xiangyang Capital acted as the long-term exclusive financial adviser. Rather than competing head-on with major tech firms for software entry points on the desktop, Violoop is building a handheld-sized external device that connects to a computer through HDMI and USB, reads screen information, and carries out actions through keyboard and mouse signals. The company says the goal is a physical personal AI agent that combines always-on context, long-term memory, local inference, and hardware-based authorization. Founder He Jialin told IT Juzi that the company wants to move beyond prompt-driven interaction and build what it calls “Artificial Intuition,” where the system prepares replies, gathers files, and organizes tasks before a user explicitly asks. Violoop says part of that depends on lifting on-device AI compute from 6 TOPS to 26 TOPS and keeping some key active interactions within one second. The company plans to launch through overseas channels in mid-September, begin a first small-scale domestic sale, and expand market coverage after October. About 12,000 overseas users have already paid deposits, the report said.

1010