Reame is a lean LLM inference server built on llama.cpp that treats cheap CPU hardware — shared vCPUs, free cloud tiers, 2-core ARM boxes — as a first-class target rather than a fallback. Its core thesis: on a CPU, never compute the same thing twice. The server targets narrow, repetitive workloads over private data where a small model matches frontier performance; the project measured 100% accuracy on long-context extraction with a 7B model on a free Oracle Cloud free tier 2-core ARM instance.

The architecture introduces several CPU-first optimizations. A persistent shared-prefix KV cache snapshots prompt prefixes to disk (zstd-compressed, checksummed, LRU-budgeted) so a system prompt is paid once and reused across restarts and processes. Palimpsest archives every completed generation into an on-disk n-gram store, drafting future requests at zero marginal cost. Il Suggeritore inverts constrained decoding: grammar structures propose format tokens (list numbering, bullets) for free on novel content. Self-regulating speculative decoding measures acceptance rates and draft economics at runtime, disabling speculation when it hurts throughput — critical on oversubscribed vCPUs where draft models run as slowly as targets. The Conclave runs --best-of N candidates in one interleaved batch, cloning the prompt KV instead of re-prefilling, and stops stragglers the moment an absolute majority agrees on the final result.

Benchmarks on an Apple M3 Pro show 3.2× speedup with speculative decoding (87% acceptance), 4.8× end-to-end from warm disk cache, 2.3× from Palimpsest on repeated requests (22→51 tok/s), and 2.1× from format drafting. The Conclave squeezed roughly one extra correct answer per quiz at ~2.5× wall time (not 5×). Two negative results shaped the design: speculation backfires on heavily oversubscribed vCPUs, and majority voting corrects variance, not bias — a 1.5B×5 Conclave lands between 1.5B and 3B, never above 3B.

Reame ships an OpenAI-compatible REST API, zero-config CLI (reame run qwen2.5-1.5b auto-configures threads, KV quantization, cache), and 210 isolated test cases. It fits document extraction, batch pipelines, privacy-bound legal/medical work, and private code autocomplete on a €5 VPS. It explicitly does not replace ChatGPT for general reasoning, agentic coding, or creative writing at scale. The engineering honesty — publishing negative results alongside wins — makes this a rare tool that admits exactly where CPU inference wins and where it doesn't.