Run free AI models on your own PC
Rosalena Harness is a desktop app that runs AI models on your machine and on free cloud tiers, through one router that picks the model and switches when one is busy. Try the window in a browser first; nothing to install, no account, no key. I'm building it in public on one RTX 3090.
Try Harness opens the app's own layout with a chat box in it, answered by a free cloud window on this site. Try Local runs a small open model on your own graphics card, with nothing sent to a server. Want to know which bigger models your PC can hold? Check your PC.
Your runs
- sum numbers.txt done
- summarize agent.mjs done
- draft release notes running
Other harnesses
- Claude Code read-only
- DSH read-only
Chat with a free open model, running in your browser
Pick a model and press Load. It downloads once into your browser's cache and then runs on your own graphics card through WebGPU, so it works offline after that. Smaller models load faster; bigger ones answer better. This is a taste of local AI, not the full harness: the desktop app runs much larger models.
Models come from the open-source WebLLM project and download from Hugging Face. Sizes shown are the graphics memory each one needs. Small models make mistakes, so check anything important.
Free models the harness runs today
These are the free entries in the harness router's model list as of October 2026. Paid models (Claude, GPT, Gemini) work too with your own API key, but you don't need one to start.
| Model | Runs on | Good for |
|---|---|---|
| Qwen3.8 27B | your GPUAbout 20 GB of graphics memory on an RTX 3090 | Writing, coding and agent tasks, fully private and free |
| DeepSeek V4.1 Flash | free cloud tierA shared free window that can answer "busy" | Fast answers when you don't have a big GPU |
| GLM-5.3 Flash | free cloud tierSame free window as DeepSeek | A second free option the router falls back to |
| Kokoro and Qwen-Image | your GPU | Voice and pictures, through the ComfyUI plugin |
One window for local and cloud models
Everything below works in the current build on my machine. What isn't built yet is on the download page.
One router
An OpenAI-compatible router on your PC. You ask for a job like "code", it picks the model, and if that model is busy or out of budget it falls back to the next one.
Local first
Your own GPU is the first choice: free, private, no per-token bill. Free cloud tiers come next, then paid models only if you add a key.
An agent that does tasks
It reads and writes files, runs tools, and hands parts of a job to subagents. Every run is logged so you can see what it did.
Pictures, video, music
A ComfyUI plugin renders images, short clips and music on the same GPU, from saved workflow presets.
MCP server
Other tools can list your runs, read them, and start tasks through MCP. It can read your Claude Code and DSH sessions too, read-only.
Voice input
Talk to it instead of typing. Speech is turned into text locally with Whisper, and it knows your project's own words so it gets function names right.
Your own browser
With the SolBrowser extension the agent reads pages in your browser with your logins, and clicks and types there, even inside embedded frames. Public pages are read without opening a browser at all.
Same model, two harnesses
I ran Rosalena and OpenCode on the same local model, Qwen3-Coder 30B on my RTX 3090, through the same llama.cpp server. Eight small coding tasks, each graded by tests the agent never saw, one run per task with a 6 minute limit. It's a small test on one machine, not a leaderboard, and I'm adding more harnesses to it.
| Harness | Tasks passed | Median time per task |
|---|---|---|
| Rosalena | 8 of 8 | 67 seconds |
| OpenCode 2.0.32 | 6 of 8 | 101 seconds |
OpenCode was faster on four of the tasks they both passed. Its two misses were a CSV conversion that came out wrong and a test-writing task where it left the planted bug in, and two of its passes ran into the time limit. Measured November 4, 2026.
Start with the free stuff
Try the window in your browser now, answered by a free cloud model. Or run a small model on your own graphics card. The desktop app is out as an early preview, version 0.1.13 for Windows and Mac. It isn't signed yet, so your computer warns you the first time you open it.