On-device AI infrastructure
Run larger, more capable AI models on the hardware you already have.
Minutus compresses large language models, including your own fine-tuned ones, so they run on a laptop, an edge device or your own servers. No retraining. Your data never leaves your infrastructure.
Qwen3-8B in 3.9 GB, as accurate as the official 4-bit release that needs 6.1 GB.
- 4.2x smaller
- Qwen3-8B from 16.4 GB to 3.90 GB View methodology
- Same score
- as the original on instruction following (82.1) View methodology
Measured on instruction-tuned models with standard task benchmarks. Methodology →
Measured trade-offs
More model per gigabyte
Instruction following (IFEval) vs. model memory, Qwen3 family
- Qwen3-1.7B
- Qwen3-4B
- Qwen3-8B
Higher and further left is better. Same evaluation harness for every model.
Your model, compressed. Not ours.
01
Any open model or private fine-tune
We compress the model your team already validated, not a fixed catalogue of pre-built models.
02
No retraining
Compression runs in hours on a single GPU instead of weeks of training.
03
Measured on your tasks
Every pilot ships with a quality report on your own workload.
Where it lands first
Also proven on Apertus-8B, the Swiss open model: 89% of benchmark quality at 3 bits, long context preserved up to 16k tokens.
Apertus results →