Underdog (Personal AI)

Founder Sigil Wen says that when he was 18, a popular email app stored his emails on its servers, and a bug leaked his private, privileged information to other users, costing him about $100,000; at the time he was living in a hacker house with Andrej KarpathyAndrej KarpathyAndrej Karpathy is an AI researcher and educator, one of the founding members of OpenAI, who served as Tesla’s Director of AI and led the Autopilot computer vision team from 2017 to 2022, founded the AI education company Eureka Labs in 2024, and joined Anthropic in 2026 to work on pre-training research; at Stanford he designed and taught the university’s first deep learning course, CS231n, and he coined the term “vibe coding.”Open the full entry. The episode motivated him to build products that help people while keeping their data safe.⁠[1][2][2]

He has been trying to put AI models on his own devices since 2018: in 2023 he got GPT-2 running on his Apple Watch, and at Airchat he worked with Naval RavikantNaval RavikantNaval Ravikant is an entrepreneur and angel investor and a co-founder of AngelList. Early on he co-founded the review site Epinions; he and Babak Nivi later grew AngelList out of the startup-fundraising blog Venture Hacks, and he co-founded the cryptocurrency hedge fund MetaStable Capital and Spearhead, which backs founders to make angel investments. His views on startups and life were compiled into the book The Almanack of Naval Ravikant.Open the full entry to put the speech recognition model Whisper on the iPhone, shipping it to 100,000 people for free. As smaller models became more capable and hardware improved, he concluded that running useful personal AI locally had become feasible.⁠[3][3][2]

Wen announced a private alpha of Underdog on X, saying he was worried about his data ending up in training sets and his emails sitting in databases, and that he did not trust the cloud because AI is very good at hacking, so he first built this locally run, on-device AI system for himself. Underdog’s official X account had been registered the day before.⁠[4][4][4][5]

The team released Husky, an inference engine built specifically for its own Woof model to run on Apple silicon. According to the team’s published tests on an Apple M5 Max, Woof 4B with the Flash draft model enabled generated up to 730 tokens per second, reaching up to 4.5 times the speed of Apple’s MLX framework on a function-edit task; Husky was faster than MLX on all 16 prompt types the team measured. The team attributed the speedup to verifying eight candidate tokens in parallel at each step and to pipelined submission of work, so that the GPU never waits on the host.⁠[6][6][6][6][6][6]

Wen published the manifesto “Own your AI, don’t let it own you,” arguing that most of the AI tokens humanity generates will soon come from the phone in your pocket and the computer on your desk, that users should not have to surrender their privacy to use AI, and that they should not pay for tokens their own computers can now generate for free. He wrote that “everything hackable will be hacked” and that the only data that cannot leak from a server is data never sent to one, so Underdog keeps all AI processing and data on the device and runs no servers for personal data.⁠[3][3][3][3][3][3]

On October 2, 2026, Wen formally announced Underdog on X, and the same day Andreessen Horowitz announced that it was leading the first funding round of its developer, Conway Research. Underdog launched as an invite-only beta, initially supporting Mac and Windows computers, with Linux, iPhone, and Android versions planned.⁠[7][2][8][8]

At the time, Underdog used a 27-billion-parameter reasoning model fine-tuned from Qwen3.8-27B, far smaller than the state-of-the-art models in data centers; Wen argued that it compares favorably in some benchmarks with Claude Opus 4.6, the top model of six months earlier, and is sufficient for everyday tasks such as shopping research and answering math homework questions. The product also encrypts the keys to the email and other accounts users authorize it to access.⁠[8][8][8][8]

The team released the open-source model Saluki 27B, which compresses the open-source model Qwen3.8-27B into a 7.89 GB 2-bit file that runs on standard llama.cpp and suits laptops with 16 GB of memory. On Underdog Bench, the team’s own 120-task tool-calling test drawn from the Berkeley Function Calling Leaderboard, it passed 88 tasks, versus 84 for the full-size model. The team also noted that a 120-task test is modest and that part of the lead is likely run-to-run variation, and it did not claim the model is smarter than the original; on competition math (AIME) it keeps only 82% to 85% of the full model’s score, and this release cannot process images.⁠[9][9][9][9][9][9][9][9][9][9]

Version 1.3.13 on October 10 let Chat generate and edit pictures locally on a Mac with 48 GB of memory or more; on-device video generation followed on October 11; and version 1.4.1 on October 13 added the ability to shop for users through Link.⁠[4][4][4][4]

Underdog calls itself “the most loyal AI” and handles email, calendar, meeting notes, dictation, and online errands. It can use the user’s browser to book flights, reserve tables, order groceries, and file insurance claims, asking before it pays; for sensitive tasks such as doctor’s appointments, prescription refills, and tax paperwork, it finds the needed information and prepares the next step for the user’s approval.⁠[10][1][10]

Underdog offers three in-house models: Woof 4B emphasizes speed at about 115 tokens per second; Saluki 27B keeps 96% of the full 27B model’s performance at about a seventh of the size; and Underdog 27B, built on Qwen 3.8 27B, scores 33.7 on the Artificial Analysis Intelligence Index versus an estimated 31.9 for Claude Opus 4.6 Max, a score that is in fact its base model’s. As of October 2026, the Mac version requires Apple silicon, the Windows version is in beta, and iPhone and iPad versions are available through the App Store and require iOS 18.⁠[10][10][10][10][10]

Underdog’s models run on the user’s computer, and email, recordings, and transcripts are not sent to cloud AI services; mail syncs directly between the user’s computer and providers such as Google or Microsoft, and mail, calendar, and saved logins are encrypted with a key that never leaves the machine, stored in the Apple Keychain on a Mac. With “Naval mode” turned on, the app talks only to the servers it needs, such as the mail provider. On the server side, only a small amount of data is kept, such as the account email, membership, the days the app was opened, and daily usage counts that can be turned off.⁠[10][11][11][11][11][11]

On September 10, 2026, Underdog passed the App Defense Alliance CASA Level 1 assessment conducted by Leviathan Security Group, meeting all 20 requirements, and the same month Google approved its restricted-scope access to Gmail and Calendar. As of October 2026, the security researchers who found holes in OpenAI were conducting a penetration test of the Mac app, the browser agent, and the website, and researcher Jane Manchun Wong was taking apart the Mac app to look for vulnerabilities; the team pledged to publish the findings and fixes when each review closes.⁠[11][11][11][11][11][11][11]

Underdog is free for users and is not ad-supported. Because the models run on users’ own devices, the developer does not pay for cloud inference; when a user asks it to buy something, it looks for the cheapest, best option and pays with a one-time card made just for that order, and the developer takes a tiny share, similar to an interchange fee. Wen says this means the product never has to mine user data and stays aligned with the user’s interests.⁠[8][8][10][8][8]

Underdog: Official website: https://underdog.ai/ (opens in a new window).⁠[5]

Conway Research: Official website: https://conway.tech/ (opens in a new window).⁠[1][12]

X: Public account: https://x.com/UnderdogAI (opens in a new window)⁠[4]