RunAnywhere (YC W26)

276 posts

RunAnywhere (YC W26) banner
RunAnywhere (YC W26)

RunAnywhere (YC W26)

@RunAnywhereAI

RunAnywhere: The default way of running on-device AI at scale. Backed by @ycombinator

Katılım Temmuz 2025
8 Takip Edilen1.4K Takipçiler
Sabitlenmiş Tweet
RunAnywhere (YC W26)
RunAnywhere (YC W26)@RunAnywhereAI·
A 27B parameter model used to need a server room. Now it runs on: • iPhone • Android • Mac @PrismML's Bonsai makes it possible. • 1-bit weights • 27B params in just 3.9GB • ~90% of full precision quality (PrismML evals) • Better than the 2-bit version at less than half the size The entire Bonsai family is now live in RunAnywhere. • 1.7B to 27B models • 1-bit + 2-bit ternary • Thinking mode On iOS and Mac, Bonsai runs through @ggml_org's llama.cpp and @Apple's MLX. On Android, Bonsai runs through @ggml_org's llama.cpp and directly on the @Qualcomm Hexagon NPU through QHexRT, our proprietary inference runtime. We built custom silicon kernels to make 1-bit inference on an NPU possible for the first time. Two years ago this needed a data center. Now it thinks in airplane mode. Now available in the RunAnywhere app on the App Store and Google Play.
English
6
18
92
10.2K
RunAnywhere (YC W26) retweetledi
Shubham Malhotra
Shubham Malhotra@ShubhamMal72313·
A 27B parameter model used to need a server room. Then it ran on your phone. Today it runs in a browser tab. The @PrismML Bonsai family is now live in @RunAnywhereAI for Web: • 1.7B → 27B models, 1-bit weights • Thinking mode • Zero install, zero servers - your prompts never leave your device Under the hood: • @ggml_org's llama.cpp compiled to WebAssembly • WebGPU acceleration when your browser has it, CPU fallback when it doesn't • Models download once into browser storage (OPFS) and stay there Same models, same stack as our iPhone, Android and Mac apps - now in the browser. Try it: runanywhere.ai/web-demo
RunAnywhere (YC W26)@RunAnywhereAI

A 27B parameter model used to need a server room. Now it runs on: • iPhone • Android • Mac @PrismML's Bonsai makes it possible. • 1-bit weights • 27B params in just 3.9GB • ~90% of full precision quality (PrismML evals) • Better than the 2-bit version at less than half the size The entire Bonsai family is now live in RunAnywhere. • 1.7B to 27B models • 1-bit + 2-bit ternary • Thinking mode On iOS and Mac, Bonsai runs through @ggml_org's llama.cpp and @Apple's MLX. On Android, Bonsai runs through @ggml_org's llama.cpp and directly on the @Qualcomm Hexagon NPU through QHexRT, our proprietary inference runtime. We built custom silicon kernels to make 1-bit inference on an NPU possible for the first time. Two years ago this needed a data center. Now it thinks in airplane mode. Now available in the RunAnywhere app on the App Store and Google Play.

English
1
2
6
286
RunAnywhere (YC W26) retweetledi
Sanchit monga
Sanchit monga@sanchitmonga22·
Now you can share benchmarks with others Here’s the 27B 1 bit model @PrismML Bonsai running on my iPhone 17 pro. Check out the new share feature in our app, to share benchmarks on your socials of any new model that comes out for @Apple mlx, @ggml_org llamacpp, @Microsoft onnx, @RunAnywhereAI QHexRT for @Qualcomm NPU or any other backend.
Sanchit monga tweet media
RunAnywhere (YC W26)@RunAnywhereAI

A 27B parameter model used to need a server room. Now it runs on: • iPhone • Android • Mac @PrismML's Bonsai makes it possible. • 1-bit weights • 27B params in just 3.9GB • ~90% of full precision quality (PrismML evals) • Better than the 2-bit version at less than half the size The entire Bonsai family is now live in RunAnywhere. • 1.7B to 27B models • 1-bit + 2-bit ternary • Thinking mode On iOS and Mac, Bonsai runs through @ggml_org's llama.cpp and @Apple's MLX. On Android, Bonsai runs through @ggml_org's llama.cpp and directly on the @Qualcomm Hexagon NPU through QHexRT, our proprietary inference runtime. We built custom silicon kernels to make 1-bit inference on an NPU possible for the first time. Two years ago this needed a data center. Now it thinks in airplane mode. Now available in the RunAnywhere app on the App Store and Google Play.

English
1
5
10
905
RunAnywhere (YC W26)
RunAnywhere (YC W26)@RunAnywhereAI·
A 27B parameter model used to need a server room. Now it runs on: • iPhone • Android • Mac @PrismML's Bonsai makes it possible. • 1-bit weights • 27B params in just 3.9GB • ~90% of full precision quality (PrismML evals) • Better than the 2-bit version at less than half the size The entire Bonsai family is now live in RunAnywhere. • 1.7B to 27B models • 1-bit + 2-bit ternary • Thinking mode On iOS and Mac, Bonsai runs through @ggml_org's llama.cpp and @Apple's MLX. On Android, Bonsai runs through @ggml_org's llama.cpp and directly on the @Qualcomm Hexagon NPU through QHexRT, our proprietary inference runtime. We built custom silicon kernels to make 1-bit inference on an NPU possible for the first time. Two years ago this needed a data center. Now it thinks in airplane mode. Now available in the RunAnywhere app on the App Store and Google Play.
English
6
18
92
10.2K
RunAnywhere (YC W26) retweetledi
Sanchit monga
Sanchit monga@sanchitmonga22·
Every API call is a reminder: You're renting intelligence, not owning it. @arthurmensch said it perfectly: if AI becomes as fundamental as electricity, nobody wants someone else controlling the switch. That's exactly why we're betting on on-device AI. When AI runs on hardware you own: - No cloud dependency. - No API outages. - No rate limits. - No surprise pricing changes. - No model disappearing overnight. - No company deciding whether your product gets to exist. If your intelligence lives behind someone else's API, your product exists at someone else's discretion. At @RunAnywhereAI, we're building toward a future where AI runs where it belongs: on your hardware, under your control.
English
1
2
4
229
RunAnywhere (YC W26) retweetledi
Shubham Malhotra
Shubham Malhotra@ShubhamMal72313·
People treat AI chats like private conversations. Cloud infrastructure can turn them into stored records. Court orders, lost legal privilege, and massive database exposures show why privacy cannot depend only on policy. @RunAnywhereAI processes AI directly on your device, reducing unnecessary cloud exposure while enabling fast, offline, private inference. Your data. Your hardware. Your control. Try on-device AI yourself. Download the RunAnywhere app from the link in the replies. #OnDeviceAI #LocalAI #RunAnywhere
Shubham Malhotra tweet media
English
1
2
4
158
RunAnywhere (YC W26)
RunAnywhere (YC W26)@RunAnywhereAI·
You can now talk to your documents completely offline. Drop your files into the RunAnywhere app and ask them anything. No cloud, no uploads, no internet needed. -> 100% on-device RAG -> PDFs, JSON & text files -> Chat across multiple documents at once -> Every answer backed by cited source chunks you can cross-check -> Rerank + multi-query retrieval for sharper answers -> Pick your own embedding + LLM models -> Your files never leave your phone Get the RunAnywhere apps now, link in the replies
English
1
7
25
1.6K
RunAnywhere (YC W26) retweetledi
Shubham Malhotra
Shubham Malhotra@ShubhamMal72313·
A camera frame goes in. A useful answer comes out. @RunAnywhereAI handles the entire vision pipeline on-device: Capture → encode image patches → combine vision and text tokens → decode the answer → report real performance metrics. No cloud inference. No API call. Your image stays local. Try RunAnywhere. Link in the replies.
English
1
2
6
144
RunAnywhere (YC W26)
RunAnywhere (YC W26)@RunAnywhereAI·
Vision models running entirely on your phone. No cloud. Point your camera at anything and AI narrates it live. Drop in a photo and chat with it. Everything on-device, your camera feed never leaves your pocket. Live now on the RunAnywhere app. Try it yourself, link in the replies.
English
3
7
12
544
RunAnywhere (YC W26) retweetledi
Shubham Malhotra
Shubham Malhotra@ShubhamMal72313·
Most speech-to-text picks an engine once and prays. Ours listens to its own doubt. Hybrid mode in @RunAnywhereAI transcribes on-device first. The local model returns two things: the transcript, and a confidence score. Above your threshold? Done. Private, offline, zero network. Below it? The SAME audio quietly retries in the cloud, and the result comes back stamped `was_fallback: true` with both confidence numbers attached. The router shows its work. Two rules make this hard to break: 1. Low battery eliminates the cloud, not the device. Sounds backwards until you remember the radio burns more power than the NPU. A dying phone should stay local. 2. Cloud engines report no confidence, so their score is `NaN`. `NaN` never triggers a cascade. Doubt flows one way: start on the device, escalate only when unsure. You send audio. You get the best transcript the situation allows, plus a receipt showing why. Try RunAnywhere yourself. Link in the replies.
English
5
2
9
247
RunAnywhere (YC W26) retweetledi
Sanchit monga
Sanchit monga@sanchitmonga22·
.@satyanadella calls it the Reverse Information Paradox. With cloud AI, you pay twice. First with money. Then with the prompts, files, corrections, evals, workflows, and memory you send to make the model useful. The more the model learns about your business, the more proprietary context you hand over to the cloud. You are not just sharing data. You are teaching another company how your business works, what makes it unique, and how you make decisions. Over time, that can weaken the advantage that keeps your business differentiated. That is why on-device AI matters. Run inference locally and your data stays with you. Latency drops. Apps can work without a constant cloud connection. Your learning loop compounds for your company. That is what we are building at RunAnywhere: hardware-native inference and open-source SDKs for running AI across phones, laptops, desktops, browsers, and edge hardware. Local first. Cloud when needed. Control always. #OnDeviceAI #EdgeAI #RunAnywhere
Sanchit monga tweet media
Satya Nadella@satyanadella

x.com/i/article/2076…

English
3
5
13
573
RunAnywhere (YC W26)
RunAnywhere (YC W26)@RunAnywhereAI·
Open weight models are everywhere. Finding the one that fits your phone and your task is the annoying part. RunAnywhere's model picker checks your device first, then shows you the models that actually fit well, tagged by what they're good at. Models that are actually good at tool calling get flagged, so you're not stuck with one that fumbles a function call. Turn on Thinking and you watch the model reason through the problem before it answers, not just a final line. No more guessing which quantized build fits your RAM, which model is solid at tool calls, or which one actually thinks. You just pick and run. Try it yourself, link in the replies.
English
1
5
15
521
RunAnywhere (YC W26) retweetledi
Shubham Malhotra
Shubham Malhotra@ShubhamMal72313·
Today, AI runs on someone else's hardware. The models are trained on data you know little about, guided by system prompts you never see, and updated without your control. As ChatGPT, Claude, and other LLMs become the default way people get information, AI sovereignty becomes increasingly important. Who controls the intelligence you rely on? As @AravSrinivas says, this will lead to the rise of local AI. That's the future we're building at @RunAnywhereAI . We're building the infrastructure for on-device AI with SDKs and inference runtimes that let models run entirely on your own hardware. Our QHexRT engine is the first to run LLMs, vision, and text-to-speech entirely on Qualcomm's Hexagon NPU. Try on-device AI yourself. Download the RunAnywhere app from the link in the comments.
English
1
2
6
283
RunAnywhere (YC W26) retweetledi
Sanchit monga
Sanchit monga@sanchitmonga22·
Also available for @Apple MLX, Sherpa-onnx @Microsoft, @huggingface @ggerganov llamacpp and @Qualcomm QHexRT NPU in our Android app. The first SDK to benchmark across multiple inference engines on ALL the devices. Our platform SDKs support: Native swift, Kotlin, react-native, flutter, Web, Windows (wip), and RCLI. The only production ready infrastructure for running on-device AI at Scale. Check it out @RunAnywhereAI
Sanchit monga tweet media
Yohei@yoheinakajima

just ran this on my phone (17 pro)

English
1
2
7
712
RunAnywhere (YC W26) retweetledi
Yohei
Yohei@yoheinakajima·
just ran this on my phone (17 pro)
Yohei tweet media
English
0
6
43
16.8K
RunAnywhere (YC W26)
RunAnywhere (YC W26)@RunAnywhereAI·
Every on-device AI benchmark you've seen was measured on someone else's phone. So we stopped posting numbers and shipped the benchmark instead. Built into the RunAnywhere app: pick any model, hit Run, and it measures performance on your device. → Tokens/sec + time-to-first-token (LLMs, vision) → Real-time factor (speech-to-text) → Generation speed (text-to-speech) → Load time + memory usage for everything Don't trust our numbers. Make your own. Download RunAnywhere. Link in the replies.
English
3
3
27
15.2K

Keşfet

@PrismML @ggml_org @Apple @Microsoft @Qualcomm @arthurmensch @satyanadella @AravSrinivas