Kept (your tasks)
â–² retained
A model that knows everything, made into one that knows your business.
Potara takes a general model and fuses, prunes, and fine-tunes it down to your work — then quantizes it to run on your own machine. Not a wrapper. A model that is actually yours.
We take a model that knows everything, and give you one that knows YOUR business — running on your own machine.
Now its own brand
Potara now lives as OpenAuryn
The fusion factory and the Auryn model have their own home and identity. This page introduces it; the product lives at openauryn.com.
The pipeline
Eight steps from a general model to a specialist that fits your GPU. The Studio drives it; you stay in control.
1
Interview blueprint
The Studio chat asks what you do, what you'll ask it, and what your work looks like. That becomes your model's blueprint.
2
Fuse merge weights
Combine multiple AI models with different skills into one set of weights. Two models' abilities, one model's size — merging doesn't add parameters.
3
Prune cut the unused
Strip out everything you'll never use. A realtor doesn't need organic chemistry. Guided by your actual work, it keeps the parts that matter to you.
4
Train repair + teach
A fine-tune on your own data that repairs the pruning and teaches it your business. This is why the result can be better at your job than the original was.
5
Quantize fit your GPU
Compressed to fit your GPU.
6
Scorecard proof
We measure it and show you: what it kept, what it lost, how much VRAM it needs, how fast it is. Proof, not promises.
7
Run it on your machine
Exports to Ollama and runs on your machine — with web search, so it can look up the facts we deliberately stripped out.
8
Re-fuse forever compounds
When a better model comes out, re-apply your profile to it. Your work compounds; you never start over.
Why the result can beat the original — at your job
It sounds backwards to make a model smaller and expect it to be better. Here's the honest reason:
Pruning removes capabilities you'll never use, which frees capacity. Training then does two things at once — it repairs what pruning disturbed, and it teaches your business on your own data. A focused specialist that has practiced your work can outperform a generalist that has to know everything. For the things it doesn't keep, step 7's web search fills the gaps.
Two real ways to run it — pick the one that fits
Both tiers run the same pipeline and produce the same quality model. The only difference is whose hardware does the math, how fast it finishes, and how big a model you can build.
"Cloud isn't for people who can't. It's for people who won't wait."
Choose BYO to save money on hardware you already own — or Cloud for speed, scale, and freedom from your own machine. Same interview, same scorecard, same export.
BYO-GPU · cheapest
Bring your own GPU
"Our intelligence, your hardware."
Best for: you own a capable NVIDIA GPU and want the lowest cost.
- The cloud plans the job — the interview, the pruning strategy, and the eval.
- Your GPU does the heavy math — merge, prune, train, quantize.
- Our cheapest tier: the compute is hardware you already own.
Minimum hardware A 12–16 GB VRAM NVIDIA GPU (RTX-class) to fine-tune a 7B model with QLoRA. Larger models need more VRAM than a consumer card has.
Training needs an NVIDIA/CUDA GPU. A CPU can't do it in a reasonable time (days–weeks vs. ~an hour), and Apple Silicon Macs can't CUDA-train — for those, use Cloud.
Cloud GPU · more power · priced by usage
Use our GPUs
"Use ours instead."
Best for: Mac users, bigger models, speed, and teams — chosen on purpose, even by people who already own a capable GPU.
- ★ On a Mac? Apple Silicon can't do CUDA training. If you're on a MacBook, Cloud is your path — no dual-boot, no workarounds. That's a huge share of professionals.
- ★ Build bigger models. A 12 GB card caps you around 7B. Cloud fuses and trains 13B, 34B, even 70B — models a consumer card simply can't hold.
- Don't tie up your machine. A fine-tune pins your GPU for hours; on Cloud you keep working.
- Finish sooner. A cloud A100 can do in ~30 minutes what a consumer card takes ~6 hours.
- Runs while your computer is off.
- Built for teams. Repeatable, auditable infrastructure — not one employee's gaming PC.
- Reliable. No thermal throttling, driver issues, or interrupted runs.
- Try before you buy a GPU. It removes the hardware objection entirely.
- No electricity cost, and no wear on your own card.
Verify it on real quantum hardware
An optional, independent integrity check — issued as a signed certificate.
+$10 add-on
Every fusion can be verified on real quantum hardware. IBM Quantum runs an independent mathematical check of your model's integrity and issues a signed certificate.
It's a check, not a boost. The quantum computer verifies your model's integrity — it does not make your model better, smaller, or faster, and it is not part of training. Think of it as a tamper-proof receipt for how your model was built: verification and provenance only.
Proof, not promises
Every fusion ships with a scorecard, so you can see exactly what you got before you rely on it. (Illustrative — your numbers come from your own eval.)
Dropped (unused)
â–¼ pruned
VRAM needed
fits your GPU
Speed
tokens/sec
Fuse a model that's actually yours.
Start with the interview. The Studio takes it from there — and when the next great model drops, you re-fuse instead of starting over.
Open Auryn Studio →