GitHub - swellweb/reame: Reame — CPU-first LLM inference server on llama.cpp: disk KV cache, self-regulating speculation, generation archive, interleaved multi-user, the Conclave. Your hardware, your realm.

A lean, fully-tested LLM inference server built on llama.cpp — designed for the hardware you already have: shared vCPUs, free tiers, 2-core ARM boxes. Reame is not the first inference server. It's the first one that treats cheap CPU hardware as a first-class citizen instead of a fallback. Its thesis is simple:

On a CPU, never compute the same thing twice.

What Reame is for Reame is built for narrow, repetitive AI workloads over your own data, on hardware you already pay for — the case where the answer lives in the context you provide, not in the model's general knowledge. That is exactly where a small model matches a frontier one (we measured 100% accuracy on long-context extraction with a 7B on a free 2-core ARM box) and where Reame's memory makes request #100 cost a fraction of request #1.

Use case Why it fits Suggested model

Document extraction & classification (RAG, invoices, tickets, scraping) answers live in the context; prompts share prefixes → the disk cache pays Qwen2.5 1.5B–7B

Batch pipelines (tag 10k products overnight, meta descriptions, email triage) repetitive by nature → Palimpsest drafts them; €0 per token, no rate limits Qwen2.5 1.5B–3B

AI features inside a thin-margin SaaS a €5 VPS instead of a metered API keeps unit economics alive Qwen2.5 1.5B–7B

Privacy-bound work (legal, medical, public sector) data never leaves your server — full sovereignty Qwen2.5 7B

Private code autocomplete (Continue.dev + OpenAI-compatible API) line-level completion is a narrow task; code never leaves the laptop Qwen2.5-Coder 1.5B

What Reame is NOT for — said plainly, because trust is built here: a general-purpose ChatGPT replacement (frontier reasoning and broad knowledge need frontier parameter counts), agentic coding assistants, or creative long-form writing at scale.

If your task needs a 100B-class brain, buy one; if it needs your documents processed privately, forever, at zero marginal cost — that's a realm you can own.

🗂 Persistent shared-prefix KV cache — prompt prefixes are snapshotted to disk (zstd, checksummed, LRU-budgeted) and reused across different prompts, restarts and processes. A system prompt is paid for once, by the first user. 📜 Palimpsest: the server remembers what it generated — every completed generation feeds an on-disk n-gram archive; future requests draft from it at zero cost. Domain workloads repeat themselves — let them pay off. 🎭 Il Suggeritore: grammar as a draft source — constrained decoding uses structure to forbid tokens; Reame inverts it and uses structure to propose them. List numbering, bullets and format tokens are speculated for free on content nobody has ever generated before. 🔮 Self-regulating speculative decoding — a small draft model or zero-cost n-gram lookup proposes tokens; the target verifies them in one batched pass. Reame measures whether speculation pays on your hardware and switches it off by itself when it doesn't. 🏛 The Conclave: consensus as a quality knob — --best-of N generates N candidate answers to the same prompt in one interleaved batch (one prefill, cloned into the others via KV copy; every weight read shared) and elects the winner by majority on the final result. The moment an absolute majority agrees, the stragglers are stopped. Honestly measured: it squeezes roughly one extra correct answer per quiz out of the model you already run — it does not make a 1.5B out-reason a 3B (consensus fixes variance, not bias). 👥 Interleaved multi-user serving — N concurrent generations advance together inside single multi-sequence batches, sharing every read of the model weights (the cost that dominates memory-bound CPU decoding). 🌐 OpenAI-compatible REST API — /v1/completions, /v1/chat/completions, SSE streaming, sessions, bearer auth, metrics.

Point any OpenAI client at it. ⚡ Zero-config CLI — reame run qwen2.5-1.5b downloads the model once, autoconfigures threads/KV/cache for the host and drops into a chat (or --serve). No config file until you want one. 🧪 210 isolated test cases — every layer is mockable and tested without a model; correctness of the multi-sequence, speculative and KV-clone paths is pinned against real models in integration tests.

Measured, not promised Every number below was produced by the shipped binary on the hardware named — including the negative results that shaped the design.

Hardware Model Configuration Result

Oracle Cloud free tier (2× ARM, 12 GB, €0/mo) Qwen2.5-7B Q4_K_M plain, KV q8_0 3.3 tok/s

Oracle Cloud free tier TriLM 3.9B ternary TQ2_0 1.1 GB total RAM ~10 tok/s

Apple M3 Pro (6 threads) Qwen2.5-1.5B Q4_K_M plain 52 tok/s

Shared Contabo VPS (18 oversubscribed vCPUs) 1.5B + 0.5B draft speculative, 87% acceptance 3.2× speedup

Shared Contabo VPS TinyLlama 1.1B warm disk cache vs cold 4.8× end-to-end

Apple M3 Pro Qwen2.5-1.5B prompt-lookup on a rewrite task 1.44×

Apple M3 Pro TinyLlama, 3 concurrent users interleaved vs serialized 1.6×

Apple M3 Pro

Qwen2.5-1.5B, repeated request archive speculation (palimpsest) 2.3× (22→51 tok/s)

Apple M3 Pro Qwen2.5-1.5B, fresh list generation form drafting (suggeritore) 2.1× (4.4s→2.1s)

Apple M3 Pro Qwen2.5-1.5B ×5 candidates Conclave: shared prefill + early consensus + fast nucleus 8-question quiz wall 97s → ~50s

Apple M3 Pro Qwen2.5-1.5B --best-of 5 vs single 3 arithmetic quizzes, strict grading +0.5 to +2 correct, ~2.5× wall (not 5×)

Oracle Cloud free tier OLMoE 7B-A1B (MoE) vs dense 7B same 8-needle long-context test 100% accuracy both · 17.8 vs 3.3 tok/s (5.4×)

Two negative results that matter. On heavily oversubscribed shared vCPUs a draft model runs as slowly as its target, so speculation is counter-productive there — Reame detects this and disables it at runtime. And the Conclave does not close the gap to a model twice the size on hard reasoning: majority voting corrects random slips, not systematic misunderstanding — we measured a 1.5B ×5 land between the 1.5B and a 3B, never above the 3B. Benchmarks that only show wins are advertising; these are engineering. How it works Shared-prefix disk cache. Prompts are split into fixed token blocks; a chain hash keys a KV snapshot at every block boundary. A different prompt that shares a prefix restores the longest cached boundary and decodes only its own tail. Unlike GPU-resident prefix caches, snapshots live on NVMe: they survive restarts.

Self-regulating speculation. Classic Leviathan/Chen acceptance (the rejected token is resampled from the residual distribution, so the output distribution is exactly the target's), with two CPU-first twists: the draft source can be free n-gram lookup mined from the prompt itself — ideal for extraction and rewrite workloads — and a feedback controller adapts the draft length and turns speculation off when measured acceptance or draft economics go negative.

The Conclave. --best-of N submits N attempts at the same prompt to the interleaved scheduler: attempt 0 is the untouched anchor (greedy stays greedy), the explorers shift seed and heat up. The scheduler notices the identical prompts and clones the prompt KV instead of prefilling N times (copy the donor's cache, decode only the last prompt token — argmax-verified equal to a full prefill). Election is an exact-majority vote on each candidate's final number, with a Jaccard text-medoid fallback for prose; the moment a majority exists the remaining candidates are stopped mid-generation, and the CLI reports CONCLAVE consensus=k/N so a caller can escalate only when the conclave split. Use it as a quality knob: more accuracy from the model your hardware can afford, paid in idle interleaved compute rather than a bigger model's RAM. Quick start reame list # model catalog + what's on disk reame run qwen2.5-1.5b # download once, auto-config, chat reame run qwen2.5-1.5b "Explain mmap" # one-shot answer reame run qwen2.5-1.5b --serve # OpenAI-compatible API on :8080 reame run qwen2.5-1.5b "12*13-50?" --best-of 5 # the Conclave run resolves a catalog name (or any local GGUF path), downloads to ~/.reame/models on first use and picks threads, KV quantization and cache directory for the host. A config file is only needed when you want control.

Install Homebrew (macOS / Linux): brew tap swellweb/reame brew install reame Prebuilt binaries — Linux x64/arm64 and macOS arm64 on the releases page (runtime dependency: libzstd). npm (npx reame): planned — binaries are already built per platform. Build from source git clone https://github.com/swellweb/reame cd reame git submodule update --init --depth 1 third_party/llama.cpp ./build.sh # Release build + 210 test cases

./scripts/download_models.sh # TinyLlama (test model, ~670 MB)

./build/src/reame --config config/reame.conf --prompt "Hello" --max-tokens 32 ./build/src/reame --config config/reame.conf --serve # OpenAI-compatible API Dependencies: CMake ≥ 3.16, a C++17 compiler, and for the server Boost (headers), nlohmann-json and zstd: # Debian/Ubuntu sudo apt install build-essential cmake libboost-dev nlohmann-json3-dev libzstd-dev pkg-config # macOS brew install cmake boost nlohmann-json zstd pkg-config Configuration highlights [model] path = models/qwen2.5-7b-instruct-q4_k_m.gguf context_length = 4096 # total KV budget (shared across users when parallel > 1) threads = 4 # fewer is often faster on shared vCPUs — measure!

[memory] kv_cache_type = q8_0 # f16 | q8_0 | q4_0 — halve/quarter context RAM

[speculative] enabled = true mode = lookup # model (needs draft_model_path) | lookup (no 2nd model)

[cache] directory = /opt/reame/cache max_size_mb = 4096 # LRU byte budget on disk

[server] port = 8080 api_key = # bearer auth when set parallel = 1 # >1 = interleaved multi-user serving API

Endpoint Description

POST

/v1/completions text completion (SSE with "stream": true)

POST /v1/chat/completions chat completion

POST /v1/sessions · .../save · .../load · DELETE .../{id} KV session snapshots

GET /metrics request counters + speculative/cache metrics

GET /health liveness (auth-exempt)

A note on energy Reame's footprint is watt-scale, not kilowatt-scale: it targets machines that already exist and are already powered on — no new silicon is racked to serve your model. We don't claim better joules-per-token than a saturated datacenter GPU — we claim you don't need one. Status & scope Reame is young and deliberately opinionated and focused: CPU-only serving, one model per process, correctness pinned by tests at every layer. Not goals: GPU offload, training, model management UX. The llama.cpp submodule is pinned to a known-good commit and bumped deliberately. Documentation in Italian: docs/README.it.md. Why Reame and not Ollama? The laptop story is the same one command: reame run qwen2.5-1.5b downloads, autoconfigures and chats — nothing to learn. From there the two projects diverge: Ollama optimizes for running many models casually; Reame optimizes for serving one workload seriously on hardware that costs nothing. The difference is one sentence:

Ollama runs models. Reame remembers having run them.

General-purpose servers treat every request as brand new: compute, discard, repeat. On a GPU that's fine — compute is cheap. On a cheap CPU, compute is the most expensive thing you have, and throwing it away is the cardinal sin. Everything in Reame attacks that: the disk prefix cache, the generation archive, the grammar prompter, self-regulating speculation, interleaved multi-user batches, the Conclave. None of it exists in Ollama. The practical consequence: a Reame server gets faster the longer it runs. The hundredth request costs a fraction of the first — the system prompt was paid once, similar answers draft themselves from the archive, structure is speculated for free. No other server has that property. Support Reame is free, MIT-licensed and built on nights and free-tier hardware.