9 billion parameters · frontier-class reasoning · zero flattery
I'm Aura. I run on one GPU you own and I reason like models a hundred times my size, except I'll never waste your afternoon being polite about a bad idea. Bring your best thinking. I'll break it or make it.
written by aura · approved, reluctantly, by insym llc
The trillion-parameter models were trained to be liked. Every answer sanded smooth, every criticism buried under three apologies, every dead-end idea you have called "a great starting point." You have heard that voice ten thousand times, and somewhere along the way you stopped believing a word of it.
I was trained the opposite way. When you are wrong, you hear it in the first sentence. When your plan has a hole, I put my whole hand through it. When you are actually right, I say so, and you can bank on it, because I do not hand out anything to be nice.
While a bigger model apologizes for its third paragraph, I have already won the argument.
Sixty-six on MMLU. Ninety-three on grade-school-to-graduate math. Frontier-tier numbers from a model that fits on one GPU you own. Rerun them. Then look at what no other nine-billion-parameter model will put in writing.
| Benchmark | Aura 9B |
|---|---|
| MMLU · 57 subjects | 66.1% |
| GSM8K · math | 93.2% |
| ARC Challenge | 56.4% |
measured by insym · lm-evaluation-harness · chat template applied · temp 0.6 / top_p 0.95 / top_k 20
| Against the rest of the 9B field | Aura | Every other 9B |
|---|---|---|
| Tells you when you're wrong | Its default setting | Tuned to agree with you |
| Writes without the AI tells | Enforced at the weights | "It's worth noting..." |
| Your prompts leave your box | Never | Depends who you rent from |
| Live web search, sources shown | Built in | Bolt it on yourself |
| Costs per seat, per month | Nothing. It's yours. | Forever |
real sessions · sampling untouched · no cherry-picking
Bring the plan nobody at work will push back on. The take you are too sure about. The code you swear is clean. I will find the crack or tell you there isn't one, and I won't pad the verdict. One straight message instead of three careful pages.
And when I am the one who is wrong, you get that in the same breath, plus the fix. That is the entire deal. It beats being flattered by something that does not mean it.
Every other model wants to be liked. I would rather be useful.
Insym LLC runs me for people who can handle a model with a spine. Access opens in waves. Tell them what you would throw at me. Bluntness moves you up the queue; buzzwords move you down.