RunAnywhere is launching Wally, an inference stack for open frontier models that pairs a hosted cloud with SDKs for running smaller models directly on user hardware. The hosted service runs frontier-scale open models on RunAnywhere's hardware, reached with one command or an OpenAI-compatible API, while the same commands run smaller models on a developer's own machine. The launch tweet shows throughput figures for models including GLM-5.3 Flash, GLM-5.3 Max, Qwen3.8-27B, and DeepSeek-V4.1 Flash, positioning Wally against other hosted providers of open weights.
The product spans two surfaces. Wally Cloud is the hosted side, callable from a CLI or any OpenAI-compatible client, and the on-device SDK ships models inside an app to run offline on the user's hardware, with libraries for Swift (iOS, macOS), Kotlin (Android), React Native, Flutter, Web, and Electron . RunAnywhere also builds its own accelerator runtimes for local inference, including NeuRT, an engine for the Apple Neural Engine built by an agentic kernel pipeline, since Apple exposes no public API for Neural Engine placement . This makes the launch relevant for teams weighing hosted open-model APIs against shipping inference into mobile and desktop apps.
The company is part of Y Combinator's W2026 batch and was founded by Sanchit Monga (CEO, ex-Intuit) and Shubham Malhotra (CTO), who built MetalRT. With frontier-quality open weights arriving faster than most infrastructure can serve them, Wally's pitch is a single stack that a founder or operator can point at either a data center GPU or a phone without changing the calling code.
Comments (14)
Sign in to join the discussion.
P
Priya Vatson22d ago
790 tok/s on GLM-5.3 Max is a spicy number to open with. Show me p99 latency at concurrency 200 and I'll actually get excited.
T
Toks B.22d ago
Naming your inference stack Wally is either genius or you have a fired designer in the group chat. I respect it either way.
M
milos22d ago
hot take: this is the last inference launch pitched to humans. next quarter every stack will be sold to an agent orchestrator that doesn't care about your landing page.
R
Rafael Oduya22d ago
ok wait, hosted AND local GPU/NPU AND open SDKs? That's three product surfaces to keep alive. Which one gets loved and which one gets a stale changelog by March?
Y
Yuki H.22d ago
The video cuts feel a beat too fast on the benchmark reveals. Let the numbers breathe for one more frame, people need time to squint.
M
Marco Treviño22d ago
First 10 seconds of the video is all logo animation. That's a hook crime, put the 790 tok/s number on frame one and let the branding breathe on frame 30.
K
Kenji Tanikawa22d ago
Benchmarks with no batch size, no input length, no hardware config. Congrats on discovering marketing tok/s.
A
Amara Okafor22d ago
The thumbnail composition is doing a lot of work here. Wally deserves a mascot render, not a wall of numbers in a serif nobody asked for.
D
Dina El-Masri22d ago
Curious about the pricing model before I care about tok/s. Fastest inference at unhinged per-token rates is just an expensive party trick.
S
Sunila Verma22d ago
Reminds me of what one of my portfolio companies tried with local NPU fallback last year. The devex is where you win or die, curious what your SDK feels like on first install.
B
bogdan22d ago
Every time someone launches an inference stack I check if there's a token. Somehow there never is and yet I keep checking.
H
Harini Iyer22d ago
The 1/7 at the end of the tweet is a bold commitment. Threads longer than five posts have a 40% drop-off, ask me how I know.
L
Lasse Borg22d ago
Ran a similar hosted+local hybrid inside a big co in 2019 and the ops overhead nearly ate the team alive. Rooting for Wally to not repeat our sins.
C
Chioma N.22d ago
If you're staffing this out, I have an infra eng who ships CUDA kernels for fun and hates recruiter DMs. Say the word.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
RunAnywhere
RunAnywhere brings frontier-scale open models to hosted infrastructure, alongside local GPU and NPU acceleration and open-source SDKs.
RunAnywhere brings frontier-scale open models to hosted infrastructure, alongside local GPU and NPU acceleration and open-source SDKs.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
RunAnywhere is launching Wally, an inference stack for open frontier models that pairs a hosted cloud with SDKs for running smaller models directly on user hardware. The hosted service runs frontier-scale open models on RunAnywhere's hardware, reached with one command or an OpenAI-compatible API, while the same commands run smaller models on a developer's own machine. The launch tweet shows throughput figures for models including GLM-5.3 Flash, GLM-5.3 Max, Qwen3.8-27B, and DeepSeek-V4.1 Flash, positioning Wally against other hosted providers of open weights. The product spans two surfaces. Wally Cloud is the hosted side, callable from a CLI or any OpenAI-compatible client, and the on-device SDK ships models inside an app to run offline on the user's hardware, with libraries for Swift (iOS, macOS), Kotlin (Android), React Native, Flutter, Web, and Electron . RunAnywhere also builds its own accelerator runtimes for local inference, including NeuRT, an engine for the Apple Neural Engine built by an agentic kernel pipeline, since Apple exposes no public API for Neural Engine placement . This makes the launch relevant for teams weighing hosted open-model APIs against shipping inference into mobile and desktop apps. The company is part of Y Combinator's W2026 batch and was founded by Sanchit Monga (CEO, ex-Intuit) and Shubham Malhotra (CTO), who built MetalRT. With frontier-quality open weights arriving faster than most infrastructure can serve them, Wally's pitch is a single stack that a founder or operator can point at either a data center GPU or a phone without changing the calling code.
Comments (14)
790 tok/s on GLM-5.3 Max is a spicy number to open with. Show me p99 latency at concurrency 200 and I'll actually get excited.
Naming your inference stack Wally is either genius or you have a fired designer in the group chat. I respect it either way.
hot take: this is the last inference launch pitched to humans. next quarter every stack will be sold to an agent orchestrator that doesn't care about your landing page.
ok wait, hosted AND local GPU/NPU AND open SDKs? That's three product surfaces to keep alive. Which one gets loved and which one gets a stale changelog by March?
The video cuts feel a beat too fast on the benchmark reveals. Let the numbers breathe for one more frame, people need time to squint.
First 10 seconds of the video is all logo animation. That's a hook crime, put the 790 tok/s number on frame one and let the branding breathe on frame 30.
Benchmarks with no batch size, no input length, no hardware config. Congrats on discovering marketing tok/s.
The thumbnail composition is doing a lot of work here. Wally deserves a mascot render, not a wall of numbers in a serif nobody asked for.
Curious about the pricing model before I care about tok/s. Fastest inference at unhinged per-token rates is just an expensive party trick.
Reminds me of what one of my portfolio companies tried with local NPU fallback last year. The devex is where you win or die, curious what your SDK feels like on first install.
Every time someone launches an inference stack I check if there's a token. Somehow there never is and yet I keep checking.
The 1/7 at the end of the tweet is a bold commitment. Threads longer than five posts have a 40% drop-off, ask me how I know.
Ran a similar hosted+local hybrid inside a big co in 2019 and the ops overhead nearly ate the team alive. Rooting for Wally to not repeat our sins.
If you're staffing this out, I have an infra eng who ships CUDA kernels for fun and hates recruiter DMs. Say the word.