🏰 Reame — ternary LLMs on ordinary CPUs

Reame is a CPU-first LLM server built around ternary models — weights that are only −1, 0 or +1. Its engine, written from scratch, serves PrismML's Bonsai-8B (2.2 GB) faster than the model vendor's own runtime: decode +7-14% and prompt reading +40-55%, measured at 2 and 4 ARM cores on the free Oracle box, 19/20 on a fact exam. Numbers, method, and the benchmarks we lose →

The chat interface below is running the exact setup described above: Bonsai-8B ternary, served by our nucleo engine, hosted on a free-tier Oracle ARM instance. Every token generated requires a real inference pass over eight billion ternary weights across four ARM cores. At roughly five tokens per second, you can expect a wait of about half a minute for longer responses.

To keep latency manageable on free hardware, the model operates in direct-answer mode—meaning there are no hidden reasoning steps consuming extra time. You can paste your own text and ask questions about it, or try querying it on general facts. It possesses the knowledge base of an 8-billion parameter model, which is substantial but not infallible; on borderline trivia, it might occasionally select the wrong fact. For full transparency, any response derived purely from the model's internal memory—rather than from context you explicitly provided—is explicitly flagged below the reply.