Introducing Tungsten: a local, private AI code editor

Most AI coding tools ask you to make a quiet trade: send your code to a server, get intelligence back. Tungsten makes the opposite bet. It is a fork of Visual Studio Code with two AI models inside the application, running on your own machine. The default build makes no outbound network request for AI at all — no account, no API key, no telemetry, no cloud.

Today it is out, it is free, and this post is the story of what it is and the one decision that shaped everything else.

The decision: put the model where the code is#

Every feature in Tungsten follows from a single principle — the same one behind ODEN, the search API: decide where the work happens, and put it in the right place. For a code editor that reads everything you write, the right place is not someone else's server. It is the machine already holding the code.

So Tungsten ships the models in the box. Two of them run as llama.cpp child processes bound to 127.0.0.1, and a build test walks the real import graph from the extension entry point and fails the build if any reachable module contains a URL to anything other than localhost. Privacy here is not a setting you trust; it is a property the build enforces.

Two models, one for each kind of help#

Coding assistance is really two jobs, and they want different models. Tungsten uses one for each.

The split is not about size. A base model like Qwen2.5-Coder is trained to continue text, which is exactly what fill-in-the-middle completion is — and it is fast. An instruct model like Granite-3B-Code-Instruct is trained to follow an instruction, which is what "explain this" or "rewrite this selection" needs. A base model cannot follow an instruction, and an instruct model is worse at fill-in-the-middle. Using the wrong one for either job is why some local editors feel sluggish and dumb at once.

What it feels like to use#

The completion model is loaded at startup and stays warm on 127.0.0.1:11435, so suggestions arrive without a cold start:

The instruct model is slower by nature — it loads on first use and has to read the surrounding types — and Tungsten is honest about that rather than hiding it. Beyond completion you get Ctrl+K inline edit, refactor presets, explain-this, generate-tests, a "suggest a better name" action, a second opinion on a diff, and a chat that guides you through the editor. Every surface that changes code shows you a diff first and re-checks the file hasn't changed since the diff was opened. The chat has no tools: it cannot read files, write files, or run commands.

No GPU required#

The bundled llama.cpp build is CPU-only on purpose. A 1.5B model is fast enough on a CPU, and running on any machine matters more than running fastest on some. The package carries a dozen CPU variants selected at runtime, so a processor without AVX2 works too.

• OS: Windows 10 or 11, x64
• RAM: 8 GB minimum, 16 GB comfortable
• Disk: ~4.5 GB (both models ship with the app)
• Network: none required
• GPU: none required

How it compares to cloud AI editors#

The honest trade-off: a local 1.5B/3B pair is not going to out-reason a frontier cloud model on a hard, sprawling task. What it gives you instead is instant, private, offline help that never sends a line of your code anywhere — and for the day-to-day of writing code, that is most of the value.

Why it isn't code-signed (yet)#

Windows will warn you that the publisher is unknown, because Tungsten is not code-signed. A Windows code-signing certificate costs several thousand kronor a year, and that money went into the product instead. SmartScreen will warn; the installer says the same thing before it installs anything and links to the source so you can check it yourself. It is a fork of open-source VS Code under the MIT licence, with Microsoft's upstream copyright intact — not affiliated with, endorsed by, or supported by Microsoft.

Get it#

Tungsten is free and stays free — it costs nothing to run because it runs on your machine, not ours.

• Download for Windows — latest release
• Read the full product page — features, architecture, system requirements
• Source on GitHub — MIT licensed

If you want the other half of what this one-person company builds, ODEN is a web search API for LLMs with the same philosophy pointed the other way: instead of moving your code to a model, it moves the web to your model in one call.

FAQ#

Is Tungsten free?#

Yes, and it stays free. There is no account, no licence key and no paid tier. It costs nothing to run because it runs on your machine. Tungsten is MIT licensed, like the Visual Studio Code source it is built from.

Does my code leave my machine?#

No. The default build makes no outbound network request for AI at all. Both models run as llama.cpp child processes bound to 127.0.0.1, and a build test fails if any reachable module contains a non-localhost URL.

Do I need a GPU or an API key?#

Neither. The bundled llama.cpp build is CPU-only, and both models ship inside the application. Autocomplete works the first time you open a file, with no network connection and no account.

Which models does Tungsten run?#

Qwen2.5-Coder-1.5B (base) handles completion, next-edit prediction and syntax fixes; IBM Granite-3B-Code-Instruct handles explain, inline edit, semantic fixes, refactors, tests and chat.

What are the system requirements?#

Windows 10 or 11 on x64, 8 GB of RAM (16 GB is more comfortable), and about 4.5 GB of free disk. No GPU and no network connection required.

Is Tungsten related to the ODEN search API?#

They are two products from the same one-person company. Tungsten needs no network at all; ODEN is a hosted web search API for LLMs. They share a philosophy, not a codebase.