Nova Project
We've spent years telling computers what to do. Recently they've started answering back.
Usually politely, and with great confidence. Not always to the point. We’re working on that.
To the point
We make open models work on the hardware you have, with the data you’d rather not send to anyone.
- 01
Language models on your own hardware
We run open models where the data lives: on Apple Silicon and NVIDIA GPUs. We match the model, the quantization and the memory budget to the actual machine, then measure what came out of it.
- 02
Our own inference engine
Written in Rust, from GPU kernels to the HTTP server, with no Python in the request path. New architectures get ported layer by layer and checked against a reference. The latest was the Mamba and attention hybrid in Nemotron-H.
- 03
Models inside apps and devices
We build computer vision models into products: one native Rust engine, ONNX Runtime and a stable API for apps on Apple and Windows, with no Python or OpenCV in tow.
- 04
Systems around the models
A model alone doesn’t get much done. We add retrieval, reranking, tool calls, JSON-schema answers and evaluation on real questions, instead of trusting a leaderboard.
MiRMiRYour models. Your machine.
A native runtime for language models on Apple Silicon (Metal) and NVIDIA GPUs (CUDA). One Rust binary with a terminal dashboard, a web console and an OpenAI-compatible server, including tool calls and JSON-schema answers.
$ cargo install mirmir
Throughput against the reference engines
The rows under 100% stay on the page too.
- Prompt processing
- Decode
- 100% = reference engine
| Model | Against reference 100%150%200% | Prompt processing | Decode |
|---|---|---|---|
| GPT-OSS-20B BF16GB10 · vLLM 0.29.0 | 199.6% | 188.0% | |
| Qwen3.8-27B BF16GB10 · vLLM 0.29.0 | 117.0% | 100.3% | |
| Qwen3.6-35B-A3B NVFP4GB10 · vLLM 0.29.0 | 102.6% | 95.4% | |
| Qwen3-4B BF16GB10 · vLLM 0.29.0 | 94.9% | 103.9% | |
| Qwen3-4B BF16M3 Max · MLX-LM | 91.8% | 80.8% |
Projects
Not everything we build answers in full sentences.
-
Skin analysis
A dermatoscope, from device firmware to apps on Apple, Windows and Android. Models assess wrinkles, pores, vessels, discoloration, structure, sebum and keratin, all computed by one native Rust engine.
-
Edward
A platform for clinics: registration, scheduling, consultations, day care, medical records, electronic signatures, a patient portal and reporting to Poland’s P1 health system, with an audit trail of operations and disclosures.
-
From Neuron to Trillion-Dollar Capex
The history and economics of artificial intelligence: twelve chapters, no mathematics, and a bill at the end of every chapter. The introduction and first chapter are free.
-
faststep
A UIKit-inspired UI framework for embedded Rust. The app describes views and behaviour; the framework takes care of touch, scrolling and redraw.
And also: a document signing service in Rust, and a field-technician app with offline work and order-workflow integration.
Write to us. We'll answer back.
Politely, with moderate confidence, and to the point.