| | --- |
| | title: README |
| | emoji: 💻 |
| | colorFrom: pink |
| | colorTo: gray |
| | sdk: static |
| | pinned: false |
| | license: mit |
| | --- |
| | |
| | # Trillim |
| |
|
| | We're building local AI that runs on the hardware you already have. |
| |
|
| | Trillim builds infrastructure for running models on consumer CPUs and edge devices — no GPU required. We train and fine-tune ternary ({-1, 0, 1}) models designed to run efficiently on commodity hardware, and build the tooling to deploy them. |
| |
|
| | ## What we believe |
| |
|
| | GPUs are powerful but expensive, power-hungry, and scarce. Ternary quantization changes the equation: models with {-1, 0, 1} weights don't need floating-point multipliers at all. The right software can make CPUs fast enough for real-time inference. AI should run anywhere — laptops, Raspberry Pis, edge devices — not just in datacenters. |
| |
|
| | ## What we're building |
| |
|
| | - **DarkNet** — our proprietary high-performance CPU inference engine purpose-built for ternary models, with hand-tuned SIMD kernels for x86 (AVX2) and ARM (NEON) - more supported architectures coming soon |
| | - **Tooling** — an OpenAI-compatible API server, CLI chat interface, LoRA adapter hot-swap, and an integrated voice pipeline (STT + TTS) |
| | - **Models** — ternary models fine-tuned and pre-quantized for efficient CPU inference, hosted here on HuggingFace. Look for the **`-TRNQ`** suffix. |
| |
|
| | ## Supported model architectures |
| |
|
| | BitNet, Llama, Qwen2, Mistral |
| |
|
| | ## Links |
| |
|
| | - [GitHub (public link will be added soon!)](https://github.com) |
| |
|
