On Oct. 2, 2026, 22-year-old developer and Conway Research founder Sigil Wen launched Underdog, a free, locally hosted personal artificial intelligence assistant engineered to challenge proprietary cloud platforms like Instinct and Muse. Backed by top Silicon Valley venture firms and prominent artificial intelligence researchers, Underdog executes its models directly on consumer hardware, ensuring user prompts, private messages, and sensitive personal files never leave the device.
The Launch: Local Intelligence with Zero Cloud Footprint
Underdog arrives amid escalating public fatigue over high monthly subscription fees and systemic data privacy compromises tied to server-hosted AI chatbots. Unlike dominant cloud assistants that send context windows filled with confidential documents, calendar entries, and meeting transcripts to remote hyperscaler data centers, Underdog is built entirely for on-device execution.
Developed under Wen’s laboratory, Conway Research, Underdog operates fully offline. The assistant is capable of drafting correspondence, indexing and summarizing local documents, parsing email archives, managing calendars, processing real-time voice dictation, and triggering autonomous desktop actions. By removing cloud data pipelines, Underdog eliminates server latency and shields user communications from server-side security vulnerabilities, compliance scraping, and commercial LLM training pipelines.
Currently distributed in an invite-only preview, the initial release targets Apple silicon Macs, with active ports under development for iPhone, Windows, Linux, Android, and dedicated Nvidia workstations.
Under the Hood: The Models and the Husky Inference Engine
Photo: dealroom.co (source)
To make on-device generation fast enough for routine desktop productivity, Conway Research co-designed proprietary model architectures alongside a dedicated execution engine named Husky.
Rather than forcing users to pull down unwieldy tens-of-gigabytes weights, Conway established a tiered model strategy led by "Woof," a compact 4-billion-parameter (4B) foundation model. Woof carries a 4-bit quantized footprint of approximately 2.37 GB, allowing it to reside comfortably inside unified system memory without bogging down everyday multitasking.
For heavy desktop reasoning, coding tasks, and multi-step orchestration, Conway engineered Underdog 27B, a 27-billion-parameter model based on a tuned checkpoint of Qwen 3.8. Conway benchmarks claim Underdog 27B outperforms previous generation tier-one frontier cloud models—specifically Anthropic’s Claude Opus 4.6—in instruction following, multi-document synthesis, and local code fixing, while acknowledging that larger datacenter clusters still hold an edge in advanced graduate-level scientific queries.
Performance centers around the Husky engine, built directly on Apple’s Metal hardware abstraction layer. In internal benchmarking across 16 core text and code evaluation suites on Apple's M5 Max chip, Conway reported Husky delivered up to 4.5 times the token-generation speed of Apple’s native MLX framework, achieving peak generation rates of up to 730 tokens per second on Apple silicon hardware.
| Model / Platform | Parameter Size | Memory Footprint | Primary Hardware Target | Peak Reported Speed |
|---|---|---|---|---|
| Woof 4B (Local) | 4 Billion | ~2.37 GB (4-bit) | Apple Silicon / Consumer Devices | Up to 730 tok/sec (Husky Engine) |
| Underdog 27B (Local) | 27 Billion | ~14–16 GB (Quantized) | M-Series Pro/Max/Ultra Macs | ~40–85 tok/sec |
| Instinct / Muse (Cloud) | Undisclosed (Frontier) | 0 GB (Cloud Hosted) | Remote Datacenter GPUs | ~60–120 tok/sec (Network Dependent) |
| Claude / ChatGPT (Cloud) | Multi-Hundred Billion+ | 0 GB (Cloud Hosted) | Cloud Clusters (H100/B200) | Variable (15–90 tok/sec) |
Underdog’s Law: Shrinking the Cloud Lag to Six Months
At the core of Conway Research is a central organizing principle that Sigil Wen terms "Underdog’s Law." The hypothesis argues that any intelligence milestone reached by flagship, multi-megawatt frontier cloud models will run natively on consumer-grade client hardware within approximately six months.
Historically, the gap between multi-thousand GPU datacenter training and consumer hardware deployment measured in years. Underdog aims to compress that window by leveraging extreme architectural sparsity, aggressive post-training quantization, specialized kernel engineering, and Apple's unified memory architecture.
Beyond basic text completion, Underdog is architected around agentic commerce and confidential inference. Because the assistant runs as an uninhibited local daemon, it can invoke system APIs, interact with local command-line tools, and orchestrate desktop workflows autonomously without triggering the security red tape and API metering that complicate cloud-hosted alternatives.
The Backstory: From Hacker House Breach to Conway Research
Photo: dealroom.co (source)
Sigil Wen’s path into personal AI infrastructure started early. Born in Toronto in 2003, Wen dropped out of the University of Pennsylvania’s Management and Technology (M&T) program at age 17 to pursue software development in San Francisco, occasionally sleeping out of a WeWork co-working space while breaking into the tech scene.
Wen lived in pioneering San Francisco hacker houses alongside researchers from OpenAI, Midjourney, and Stability AI, and became an early tester of Anthropic’s initial Claude architectures. He later teamed up with AngelList founder Naval Ravikant to co-found and engineer the voice-first platform Airchat, went on to secure a 2025 Thiel Fellowship, and founded Extraordinary.com, an immigration platform assisting global technical talent in securing US O-1 and EB-1A visas.
The pivot toward private local computing was sparked by a concrete security failure. While sharing workspace with former Tesla and OpenAI scientist Andrej Karpathy, a data leak in a cloud-based email application inadvertently exposed Wen’s sensitive personal emails to external users. The event shaped Wen’s belief that relying on remote multi-tenant servers to protect intimate personal data is structurally flawed: if personal data lives on an outside server, it will eventually leak.
How Underdog Compares to Instinct, Muse, and Cloud Giants
Modern enterprise assistants like Instinct, Muse, Microsoft Copilot, and Google Gemini run inside cloud ecosystems. While these platforms tap into massive compute clusters, they present clear operational drawbacks for privacy-conscious users:
- Zero Cloud Dependency: Instinct and Muse route queries through external cloud servers, creating persistent threat surfaces. Underdog never transmits user prompts or metadata over the wire.
- Free Operations: Because the user provides the electrical power and compute via their laptop or desktop chip, Conway does not shoulder inference costs, allowing Underdog to remain free without recurring subscriptions or ad-supported data harvesting.
- Completely Offline Support: Cloud assistants drop connection the moment a device goes offline or joins a restricted enterprise firewall. Underdog functions without degradation on an airplane or off-grid.
- System-Level Context: Traditional cloud tools see only what users paste into a chatbox or connect through cumbersome cloud connectors. Running locally gives Underdog low-overhead access to indexed hard drives, terminal shells, and local system notifications.
Cloud solutions retain an advantage in world knowledge breadth, encyclopedic indexing of real-time web news, and specialized reasoning across deep mathematical domains. However, for routine productivity—such as clearing inbox backlogs, triaging schedules, drafting documents, and scanning proprietary PDFs—Underdog’s local footprint matches or exceeds user needs without the accompanying privacy risks.
Venture Backing: Silicon Valley’s Elite Rally Behind Local AI
Photo: ababnews.com (source)
Despite entering a crowded market dominated by well-funded tech conglomerates, Conway Research has secured heavyweight backing. Silicon Valley venture capital powerhouse Andreessen Horowitz (a16z) led the company’s unannounced seed funding round, with partner Chris Dixon championing Wen’s push toward data sovereignty.
Institutional backers include Vinod Khosla’s Khosla Ventures, Hummingbird VC, SV Angel, Breakneck Ventures, and Anthology—an investment vehicle jointly operated by Anthropic and Menlo Ventures.
A notable roster of prominent angel investors and engineers also joined the round:
- Patrick Collison (CEO and Co-Founder of Stripe)
- Naval Ravikant (Co-Founder of AngelList)
- Guillermo Rauch (CEO of Vercel)
- Noam Brown (Reasoning Researcher at OpenAI)
- Thomas Wolf (Co-Founder and Chief Scientist at Hugging Face)
- Logan Kilpatrick (Developer Ecosystem Leader)
- Charles Songhurst (Prolific Tech Investor and Former Microsoft Strategist)
What This Means for Users and Hardware Requirements
For consumers and knowledge workers, Underdog marks a shift away from closed, cloud-metered AI toward self-contained personal appliances. Legal professionals, healthcare administrators, corporate executives, and software developers who are legally barred from pasting proprietary code or private records into public cloud models can deploy local models without risking regulatory compliance breaches.
Running high-throughput models locally does come with strict hardware demands:
- Apple Silicon Focus: The initial build requires macOS running on Apple silicon (M1 chips through M5 generations). While the lightweight Woof model runs comfortably on standard 8 GB and 16 GB configurations, running Underdog 27B smoothly requires at least 32 GB of unified memory.
- Battery Impact: While Husky delivers optimized performance per watt compared to unoptimized runtimes, continuous local inference will drain laptop batteries faster than passive browser-based web chats.
What Happens Next: The Road Ahead and Unanswered Questions
Conway Research plans to gradually expand its invite queue over the coming months while finalizing cross-platform builds for Windows Copilot+ PCs, Linux platforms, and mobile devices.
Significant practical questions remain. While running local open models protects day-to-day user prompts, Conway has not yet laid out its long-term commercial monetization plan. Supplying consumer software for free avoids the monthly server bills that strain cloud AI companies, but maintaining model training clusters, rolling out security updates, and engineering custom kernels will require continuous capital.
Furthermore, the long-term validity of Underdog’s Law faces an inevitable physics test. While compressing frontier capabilities into consumer memory was feasible when transitioning from early generation LLMs to mid-tier foundation models, the gap may widen if future frontier architectures grow dramatically in scale or rely on continuous massive inference compute clusters. Whether 2.4 GB to 15 GB local models can match those frontiers remains the core technological question Conway must prove over the coming year.
FAQ
What is Underdog AI?
Underdog is an on-device personal AI assistant developed by Sigil Wen’s Conway Research. It processes emails, schedules, local files, and coding tasks entirely on your local computer without sending personal data to cloud servers.
Is Underdog completely free to use?
Yes, Underdog is free for consumers. Because inference runs locally on the user's personal hardware rather than centralized cloud servers, the platform eliminates the recurring hosting expenses that force competitors to charge monthly subscriptions.
What devices can currently run Underdog?
Underdog is currently available in a closed preview for Apple silicon Mac computers. Versions for iPhones, Windows, Linux, Android, and Nvidia-powered hardware are actively in development.
How does Underdog perform compared to cloud models like Claude or GPT?
Its compact Woof (4B) model is designed for everyday tasks, while its Underdog 27B model matches or surpasses previous-generation frontier models like Claude Opus 4.6 in instruction following and code editing. However, massive cloud clusters still hold an advantage in deep academic, scientific, and specialized knowledge tasks.
Who funded Underdog and Conway Research?
The company is backed by Andreessen Horowitz (a16z), Khosla Ventures, Hummingbird VC, Anthology (Anthropic/Menlo Ventures), and angel investors including Patrick Collison, Naval Ravikant, and Guillermo Rauch.
