RunAnywhere Wally Adds Seven Open Models: What the CLI Does and What Is Unverified

RunAnywhere says seven more open models are now live in Wally, its command-line tool for pointing coding agents at open models. According to the company's X account, the new additions are Kimi K3, DeepSeek V4 Pro (Fast), GLM-5.3 (Fast), MiniMax M3, DeepSeek V4.1 Flash (Fast), DeepSeek V4.1 Flash 1017 and DeepSeek V4 Flash 0731. An aggregator that picked up the post counts ten models in the catalog, with Kimi K3.1 rumored next; we could not confirm the total of ten or the Kimi K3.1 rumor on RunAnywhere's own pages.

The more durable story is the tool. Wally is a MIT-licensed repository that does one practical thing: it lets you run a coding agent you already like on an open model, hosted or local, with a single command. This post explains how it works, how to try it, and which numbers come only from the company.

TL;DR

What does Wally actually do?

The README describes Wally as "a command-line tool that connects your coding agent to the model of your choice." With one command it opens an agent already configured for the model and installs the agent first if it is missing. You can use a hosted model such as GLM 5.3 Flash, billed against RunAnywhere credit, or download an open model and run it locally without an account. It can also chat with a model in the terminal and serve it through an OpenAI-compatible API, as the RunAnywhere documentation also describes.

A laptop outline with a chip, illustrating how the Wally CLI can run an open model locally on Apple Silicon

The quick start in the README is short:

curl -fsSL https://raw.githubusercontent.com/RunanywhereAI/wally/main/install.sh | sh
wally account login
wally opencode -m glm-5.3-flash

The default model is glm-5.3-flash, and wally models default <model> changes it. On Windows the install uses a PowerShell script. As with any installer you pipe into a shell, read the script first. The README says macOS support is Apple Silicon only and that Linux needs glibc 2.39 or newer, so Ubuntu 24.04 and Debian 13 qualify while older distributions do not.

Why does this matter if you already use a coding agent?

Most coding agents were built around one vendor's model. Swapping in an open model usually means hand-editing config, base URLs and keys, and testing whether tool calls survive the swap. Wally packages that wiring. The pitch is that you keep your agent and change the engine underneath.

That fits a trend we have tracked. Open models are now cheap and good enough that many teams route part of their traffic to them. Our small models comparison lines up the cheap tier, and the DeepSeek V4.1 Flash 1017 write-up covers one of the models in this catalog. OpenCode's own free tier is covered in our OpenCode DeepSeek V4.1 Flash note. Wally is a different route to the same place: a single account in front of several open models.

The seven models, in context

A bridge between two landmasses with dots crossing, representing open models from Chinese labs reaching developers through one CLI

"Fast" in the list is RunAnywhere's label for its faster serving tier. The company has not published, in anything we read, how that tier differs technically from a standard deployment, so treat it as a product name.

What is claimed and what is verified

The most checkable facts are the repository's: the MIT license, the commands, the supported platforms and the roughly 1,600 stars and 439 commits we saw. Those are visible to anyone.

The performance claims are not independently verified. A launch article on RuntimeWire reported that RunAnywhere claims 380 tokens per second on GLM-5.3 Flash and noted that the thread omits the prompt, hardware and concurrency settings needed to reproduce it. Compare against independent provider benchmarks. We did not run our own test, and RunAnywhere's own post is the origin of the number.

Other things the sources do not establish:

• Hosted pricing. The README says hosted models bill against credit but does not list per-token prices in the page we read.
• Which "Fast" tier maps to which hardware.
• Data handling for hosted requests. RunAnywhere has terms of service and a privacy policy; read them before sending proprietary code to any hosted endpoint.
• Catalog size. Seven additions are confirmed by the company's post; the total of ten comes from an aggregator.

The local path, and the Eve decision model

For privacy-minded developers the local option is the more interesting one. The README says a local model needs no account. The project certifies Qwen3 4B Instruct as its local model for coding agents, tested for tool calls and long context, per the repository. A 4B model will not match a hosted frontier-sized one on hard tasks, so expect to use local mode for small edits, private code, or offline work.

The README also introduces Eve, RunAnywhere's hosted decision model. Rather than writing text it scores explicit labels:

wally decisions --input "Checkout is blank after Pay" \
  --ask "Is this a software bug?" \
  --choice "Owner=frontend,payments,account"

You can pull a local checkpoint, clef-flash-9b, and run the same command with --local. This is the same family of "decision model" ideas we covered in Ollaya, an Ollama-style runtime for decision models and llama.cpp's decision model support. It is useful for routing, triage and classification where you want a label, not a paragraph.

How does Wally compare with other ways to run open models?

The honest summary: Wally is convenience tooling. It does not make any of these models better; it lowers the cost of trying them inside an agent you already use.

Where Wally came from

RunAnywhere launched Wally on September 21, 2026, according to RuntimeWire's launch coverage, which cites the company's X post as its primary source. RuntimeWire describes Wally as an inference stack that runs open models locally, on RunAnywhere's cloud, or on customer-controlled infrastructure. The October catalog expansion is the first sign of how quickly the hosted side is growing.

How to try it safely

1. Read install.sh before running it, or build from source using the Rust and CMake instructions in CONTRIBUTING.md.
2. Start with the local Qwen3 4B Instruct path if you only want to test the workflow with no account.
3. For hosted models, use a throwaway repository first and check what the agent sends.
4. Compare output quality against your current model on three or four of your own tasks before switching anything.
5. If speed matters, measure tokens per second yourself instead of relying on the 380 figure.

What to watch next

• Whether RunAnywhere publishes hosted pricing and a model-by-model catalog page.
• Independent provider benchmarks for the Fast tier.
• Whether Kimi K3.1 or another new open model lands in the catalog quickly, which would show how fast the company can add models.
• Whether coding-agent vendors respond to tools that redirect their agents to open models.

Bottom line

Wally is a small, MIT-licensed tool that makes it easy to run a coding agent on an open model, with seven new models announced in the catalog this week. The code and install path are real and public; the speed and catalog-size claims are the company's or an aggregator's and are not independently confirmed. If you like the idea, test it on your own tasks.

Related reading

• Small models compared: Haiku 5.5, GPT-6 Luna, DeepSeek V4.1 Flash, Gemini 3.7 Flash, GLM 5.3 Flash
• DeepSeek V4.1 Flash 1017: open weights, MIT, benchmarks
• MiniMax M3.1 Flash: pricing, weights and license
• Kimi K3 open weights
• Prime Inference launch
• DwarfStar DS4 local inference
• Ollaya: Ollama for decision models

Details are accurate as of October 23, 2026, based on the Wally repository README, RunAnywhere's X posts as relayed by coverage, and a RuntimeWire launch article; prices and the catalog may change.