On a Monday in early March, as most of Silicon Valley was still digesting the weekend, a small startup called RunAnywhere flipped a switch. Their software development kit went live—open-source, no less—with a provocative pitch: mobile developers could now run AI models directly on phones and tablets, bypassing the cloud altogether. No API keys. No per-request charges tallied by some hyperscaler. Just code running on the device in a user's hand.
The company is two people. Both came through Y Combinator's Winter 2025 batch. And within days of their March 3 launch, their GitHub repository had racked up more than 10,000 stars, a velocity that raised eyebrows even in a community accustomed to hype cycles.
Whether RunAnywhere—co-founded by Sanchit Monga and Shubham Malhotra—represents a genuine shift in how enterprises think about AI infrastructure, or simply another well-marketed toolkit in an increasingly crowded field, remains an open question. But the timing of their arrival is hard to ignore. Apple, Google, and Meta are all racing to bake AI directly into operating systems, and that's creating both opportunity and chaos for anyone trying to build on top.
The Fragmentation Problem
Here's the bind developers face: on-device AI is the future, or so the platform vendors keep insisting. Apple has Core ML and Neural Engine optimizations. Google pushes TensorFlow Lite (recently rebranded as LiteRT). Meta just brought ExecuTorch to general availability last fall. Each framework is finely tuned for its own ecosystem. None of them play particularly nice together.
RunAnywhere's bet is that enterprises don't want to maintain separate codebases for iOS, Android, React Native, Flutter, and the web. They want one SDK that abstracts away the platform-specific gymnastics. More than that—and this is where the founders think they've found their wedge—they want a control plane that lets them manage AI deployments the way they manage any other software: pushing updates, monitoring usage, routing traffic based on policies.
The SDK handles multimodal workloads: large language models, speech-to-text, text-to-speech, computer vision. Once a model is downloaded to a device, everything runs locally. No network calls. The company's documentation, reviewed in early March, is explicit: "all AI inference runs 100% on-device." You could be in airplane mode and the thing would still work.
Under the hood, RunAnywhere wraps existing inference engines—llama.cpp for language models, Sherpa-ONNX for speech tasks. It supports model families like Llama, Mistral, Qwen, and SmolLM. Native Swift packages target the full Apple hardware stack (iOS, macOS, tvOS, even watchOS, though you'd need iOS 17 or later and Apple Silicon on the Mac side). Kotlin handles Android. Cross-platform developers get dedicated packages for React Native, Flutter, and web browsers.
The Control Plane Angle
What distinguishes RunAnywhere from, say, just cloning llama.cpp and figuring it out yourself? The control plane. That's the centralized dashboard where enterprises can push over-the-air model updates using differential packaging—so you're not shipping multi-gigabyte files every time weights change—and route workloads between on-device and cloud inference based on rules. Maybe a user's phone is too old, or the battery is low, or the network connection is spotty. In those cases, the system can fall back to a cloud API.
This matters more at scale than it does for a single app. If you're managing AI across thousands or millions of endpoints, visibility becomes critical. Which models are running where? How often? Are updates propagating? RunAnywhere's March 3 press release framed this as the difference between a hobbyist project and enterprise infrastructure: companies need "structure, visibility, and control."
The technical architecture showed up in GitHub release v0.17.5, tagged on January 26. The SDK uses llama.cpp as its LLM backend, Sherpa-ONNX for Whisper-based speech recognition and Piper neural text-to-speech, and includes voice activity detection for stringing together full voice assistant pipelines. On Apple devices, it taps Core ML and Metal for acceleration. Android gets JNI optimizations.
A mid-February blog post laid out what the founders call their "unified runtime" strategy: developers write against a single API, and the SDK handles platform-specific acceleration under the hood. You can also chain models—speech-to-text feeding into an LLM feeding into text-to-speech—all executing locally without a single network round-trip.
Demo apps exist on both the App Store (updated as recently as late February) and Google Play (refreshed in early March). The iOS version has a perfect 5.0 rating, though from just three reviews. The Android app shows "100+ downloads" as of early March, which is... a start.
The Economics, and the Asterisks
RunAnywhere's core value proposition is cost. In a mid-January blog post announcing their YC launch, the founders claimed developers could "cut cloud inference costs by up to 90%" by running models locally. That's a vendor claim, not an independent benchmark, but the arithmetic is straightforward: if inference happens on the user's device, you're not paying AWS or Azure per API call. The cost shifts to device battery, storage for model weights, and the operational complexity of managing software across a fragmented fleet.
Public pricing for the control plane isn't listed anywhere. The company's website directs prospects to "Book a Demo," which is standard for enterprise SaaS but also means we have no idea what they're charging. No reference customers are named in the press release or on the website as of early March.
And then there's the traction narrative. The GitHub stars—10,200 as of March 5—are impressive in raw numbers. The company's YC Launch page, posted roughly a month earlier, showed around 3,900 stars. That's growth of more than 6,000 stars in a few weeks, which is either genuine viral adoption or something else.
Something else did come up. A "Tell HN" thread on Hacker News in late February surfaced allegations of "spammy outreach" and questioned the star patterns. Users noted recently registered domains tangentially related to the project. One of those was OpenClawPi.com, which as of early March advertised a Raspberry Pi-based personal AI device "Powered by RunAnywhere SDKs," complete with a coupon code "RunAnywhereYC26" and "Backed by Y Combinator" branding. No formal relationship between OpenClawPi and RunAnywhere is confirmed on RunAnywhere's official channels, and the affiliation—if any—remains murky.
The company is still a two-person team, according to the Y Combinator directory snapshot from late February. Diana Hu is listed as the primary YC partner. An SEC Form D filing dated October 16, 2025, shows a $10,000 amount, which could be a founder setup or test filing rather than a meaningful funding round. In other words: this is very early.
A Crowded Space, Different Angles
RunAnywhere is hardly alone in chasing on-device AI. Meta's ExecuTorch went GA last October and now powers on-device workloads across Meta's own apps. Google announced LiteRT in late February as the successor to TensorFlow Lite, positioning it as a "universal framework" with WebGPU support. NimbleEdge open-sourced its DeliteAI agentic platform last summer. Liquid AI released its LEAP dev kit around the same time.
What RunAnywhere offers that lower-level runtimes don't—at least in theory—is the abstraction layer and the fleet management. Developers don't need to wrangle CoreML optimizations separately from ONNX runtimes; they call the same SDK methods. The control plane adds operational tooling that matters more at enterprise scale than for a solo developer hacking on a side project.
The company has been active in the developer community. They sponsored hackathons in February, including events at the University of Waterloo and MinneHack. They got some attention with a Show HN post for an on-device browser agent project that runs AI models in Chrome using WebGPU. A Japanese tech outlet, GIGAZINE, covered that browser demo in mid-February.
The SDK is fully open-source under the Apache-2.0 license, which likely accelerates adoption—or at least GitHub stars. Open source also means the code is auditable, which matters if you're running sensitive inference workloads locally on user devices.
The Harder Questions
RunAnywhere is making a clear bet: AI runs locally by default in the future, and cloud inference becomes the exception rather than the rule. Co-founder Shubham Malhotra told the press his platform gives companies the "structure, visibility, and control" to operate AI across device fleets at scale.
Whether enterprises actually buy that depends on things the company hasn't yet proven publicly. Independent benchmarks of cost savings and performance would help. So would reference customers willing to go on record. The control plane needs to demonstrate robustness under production workloads, not just demo apps with three App Store reviews.
The SDK is out there. The documentation is live. Developers can clone the repo today and start building. But the harder question—whether on-device AI at enterprise scale is an operational problem companies will pay to solve, or just another open-source toolkit in a drawer full of them—will take more than a few weeks of GitHub stars to answer.
For now, RunAnywhere is in the game, actively shipping, and betting big on local-first AI. With Y Combinator backing and at least some developer attention, they've got a window to prove the thesis. Whether that window stays open long enough for two people to build an enterprise business? That's the question they're racing to answer.
