Ternary Precision
-1, 0, +1 weights reduce memory, bandwidth and compute while preserving quality.
Lighter inference.
Sharper model path.
Real bundle.
RPT packs, optimizes and runs LLMs with ternary precision. Faster inference. Smaller models. Real world performance.
Public research is open.
Scripts, notebooks, results and runtime experiments are available.
Fork it, reproduce it, improve it.
Scripts, notebooks, results and runtime experiments are available.
Fork it, reproduce it, improve it.
Packed Models
High density packing shrinks models up to 50% vs standard formats.
Faster Inference
Optimized kernels deliver real speedups across consumer hardware.
Production Ready
Stable runtime, simple APIs and tooling for real world deployments.
Benchmarks
Accuracy (MMLU)
100/100
Custom Benchmark
Size Reduction
~50%
vs Original FP16
Speedup
2.4x
Avg. Inference
Memory Savings
48%
Peak Usage
Gemma 2B Ternary vs FP16 - RTX 3060 12GB - Batch Size: 1 - Seq Len: 1024View full report ->
Runtime
Engineered to run.
Built to scale.
RPT is a lean, high performance runtime path for ternary packed models. The public repo keeps the notebooks, loaders and result trail visible.
- vValidated Linux/NVIDIA target
- vOpen notebooks and scripts
- vStreaming & batching
- vPublished result trail
Open Source
Build from the working trail
The useful pieces are public: scripts, notebooks, runtime loaders, result files, manifests and model notices. Use them to reproduce the path or build a new one.
- vResearch notebooks
- vRuntime loaders
- vBenchmark JSON/CSV
- vModel license notices