Forge 20B demo is live
Write code
offline.
VYRON Forge 20B is an open coding assistant that runs on your own machine. It answers in code first and never phones home. No sign-up, no API key.
- Runs with
- Ollama
- llama.cpp
- LM Studio
- VS Code
Announced at launch
- ~13 GBdownload, once
- 16 GBfree memory needed
- 0accounts or keys
- 3desktop OSes
01 / Playground
Ask small. Get code back.
Forge is tuned to reply with the snippet first and keep the talk to one line. These are recorded sample answers. Forge can still be wrong, so check what it gives you.
02 / Install
From zero to first answer
in about five minutes.
Forge runs through Ollama on Windows, macOS and Linux. On VS Code, add it as a local model in an extension like Continue.
-
1
Install Ollama
$brew install ollama -
2
Pull the model
Forge 20B runs on OpenAI's open-weight gpt-oss-20b. About 13 GB, downloaded once.
$ollama pull MODEL -
3
Ask it something
Leave off the prompt to open a chat. For Forge's code-first style, use the Modelfile from Hugging Face.
$ollama run MODEL "Dedupe a list in Python"
03 / Hardware check
Will my laptop handle it?
Set your memory and chip. This is a rough estimate for the MXFP4 build (about 13 GB). We'll publish measured speeds once we've benchmarked them.
Chip
- Forge + context
- OS & apps (example)
- Headroom
04 / Roadmap
A family of specialists,
shipped one at a time.
Each model is narrow on purpose. It starts from the open base that suits its job and gets its own name.
-
Demo
Forge
Writes, fixes and explains code across popular languages. Runs today as a 20B demo on gpt-oss-20b.
vyron-forge-20b -
Building
Lens
Reads a transaction or contract and explains what it actually does, in plain words.
vyron-lens -
Planned
Lingo
Localizes docs, UI strings and wallet prompts while keeping technical terms consistent.
vyron-lingo -
Research
Pilot
An agent that reads a repo, makes the smallest change that works, runs the tests and stops.
vyron-pilot
05 / Spec sheet
No mystery meat.
Same facts as the model card. If something changes, it changes in both places.
- Model
- VYRON Forge 20B (demo)
- Base
- OpenAI gpt-oss-20b, ~21B parameters (mixture-of-experts)
- Tuning
- VYRON system prompt and settings, no fine-tune yet
- Data
- No extra training data yet (base model by OpenAI)
- Format
- GGUF, MXFP4, ~12 GB
- Runtimes
- Ollama, llama.cpp, LM Studio, node-llama-cpp
- License
- Apache 2.0
06 / FAQ
Quick answers.
Something missing? Ask us on X or Telegram.
What is VYRON?
An open project building small, single-purpose AI models that run on ordinary hardware. Forge, our coding model, is the first one out.
Which model is Forge built on?
Forge 20B runs OpenAI's open-weight gpt-oss-20b with VYRON's system prompt and settings. There's no VYRON fine-tune yet, so the quality you get is gpt-oss-20b's.
Does anything get sent to a server?
Not when you run Forge yourself. Once the weights are downloaded it runs fully offline, and your prompts and code stay on your machine. The live demo on this website is the exception: to answer without an install, it sends your prompt to Cloudflare Workers AI. VYRON doesn't store demo prompts.
How good is it, honestly?
It handles everyday snippets, fixes and boilerplate well for a model that runs on your own machine. It's not a replacement for a frontier model on big refactors or tricky reasoning. Treat its output like a fast first draft.
Should I deploy smart contracts it writes?
Not without a review. Use Forge to draft and learn, then test thoroughly and get an audit before anything touches real funds.
Is it free to use commercially?
Yes. The weights are released under Apache 2.0, so you can use, modify and ship them, including in commercial products.
What's the $VYRON token?
The community token around the project. The official contract address is the one shown at the top of this page. Copy it from here and ignore addresses posted anywhere else.
Your prompt.
Your hardware.
The model card, system prompt and Ollama Modelfile are on Hugging Face.