The YC W26 company is extending its on-device stack into hosted inference, claiming 380 tokens per second on GLM-5.3 Flash.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [RunAnywhere on X](https://x.com/RunAnywhereAI/status/2102145104450601419)
Why it matters #
RunAnywhere is turning an on-device runtime into a hybrid inference business, using coding agents to sell hosted GPU capacity without abandoning local execution.
Sanchit Monga (@sanchitmonga22) and Shubham Malhotra launched Wally on September 21st, an inference stack for running open models locally, on RunAnywhere's cloud, or on infrastructure controlled by the customer.
https://x.com/RunAnywhereAI/status/2102145104450601419 The release pushes RunAnywhere beyond the on-device AI pitch that got the five-person San Francisco company into Y Combinator's Winter 2026 batch. Monga previously built mobile SDKs at Intuit, while Malhotra worked on infrastructure at Microsoft Azure and Amazon Web Services. Their original product focused on routing AI workloads across phones and computers without forcing developers to manage separate runtimes for each chip.
Wally extends that routing model to open models too large for a laptop. In a seven-post thread on X, RunAnywhere described hosted access as an explicit option rather than an automatic fallback. Developers can also deploy Wally using bring-your-own-cloud or on-premises infrastructure, according to RunAnywhere.
The open-source Wally command-line tool connects models to Claude Code, Claude Desktop, OpenCode, Hermes and OpenClaw. It can also serve downloaded models through an OpenAI-compatible local API. The MIT-licensed repository had about 1,500 GitHub stars and 201 commits when reviewed.
RunAnywhere's speed claim has a narrower benchmark behind it
RunAnywhere says Wally generated 380 tokens per second with GLM-5.3 Flash, 790 tokens per second with GLM-5.3 Max, 485 tokens per second with Qwen3.8-27B and 615 tokens per second with DeepSeek-V4.1 Flash. RunAnywhere also claims time to first token below 300 milliseconds.
In a test using the same prompt and OpenCode harness, RunAnywhere says Wally delivered 43% more throughput than Nebius, 3.1 times Fireworks AI's throughput and 4.5 times the model developer Z.ai's throughput. RunAnywhere says it repeated the race three times, although the thread does not publish the full prompt, request logs, hardware configuration or concurrency settings needed to reproduce those results independently.
RunAnywhere separately compared its 380-token result for GLM-5.3 Flash with September figures from Artificial Analysis. The company's chart listed Nebius at 313 tokens per second, Fireworks at 185 and Z.ai at 72.
The broader claim that Wally is the fastest place to run open frontier models remains unproven by the public data. Artificial Analysis' current provider table does not list Wally among the 17 measured GLM-5.3 Flash providers. It places Inco's fast endpoint at roughly 480 tokens per second and Nebius at about 356, with results that can change as providers adjust capacity and serving software. RunAnywhere's head-to-head test did not include Inco.
That distinction matters because inference benchmarks are sensitive to prompt length, output length, batching, quantization, hardware allocation and traffic conditions. A provider can lead a controlled single-request test without delivering the same advantage under sustained production load.
Wally is launching as a preview
RunAnywhere's terms for Wally classify the hosted service as a preview without uptime or support-response commitments. Accounts receive $5 in promotional credit and then use prepaid, per-token billing. RunAnywhere has not attached per-model prices to the launch thread.
RunAnywhere's privacy policy says production Wally requests use zero data retention for prompt and completion text. RunAnywhere retains separate usage and billing records, including token counts, timestamps, model identifiers, request IDs and spending. The policy identifies Google Cloud as the processor serving Wally's hosted models.
The public repository documentation currently names GLM-5.3 Flash, Qwen3.8-27B and Gemma 4 as hosted models. The launch benchmark also includes GLM-5.3 Max and DeepSeek-V4.1 Flash, indicating that the service catalog is moving faster than its README.
Wally gives RunAnywhere a paid cloud path alongside its open-source local runtimes. Coding agents are the initial distribution wedge because their long outputs make inference throughput immediately visible, while compatibility with existing harnesses lowers the cost of switching providers. RunAnywhere can use the same command-line interface to keep small workloads on a developer's machine and sell credits when a larger model or job requires hosted GPUs.
That hybrid model is the practical version of RunAnywhere's original local-AI thesis. Developers choose where each request runs, while RunAnywhere controls the interface, routing layer and billing relationship across both sides.