Become a sponsor to Marco Caciotti
I build open-source tools that squeeze real value out of hardware you already own.
Reame — a CPU-first LLM inference server built on llama.cpp, with one thesis: never compute the same thing twice. Disk-persistent KV cache that survives restarts, self-regulating speculative decoding, a generation archive that drafts future answers, and the Conclave (consensus as a quality knob). It serves a 7B-class MoE at 17.8 tok/s on a free 2-core ARM box — every benchmark in the README is reproducible, negative results included.
Your sponsorship directly funds the Reame roadmap:
- ARCA — a shared memory daemon ("Redis for AI"): responses, KV blocks and draft corpus shared across your whole fleet
- warm-ahead prefill — documents digested in the background, before anyone asks
- first-class MoE serving on cheap CPUs
TargetBridge — reuse old iMacs as displays and build multi-Mac workflows over Thunderbolt. Sponsorship keeps it maintained, tested and compatible across Macs and setups.
Practical software, honestly benchmarked, for hardware that already exists. If these tools save you API bills, GPU rent or a trip to the recycler, sponsoring is the best way to keep them alive.
Current sponsors 2
Past sponsors 1
$3 a month
SelectSupport ongoing development of Reame and TargetBridge — bug fixes, compatibility work and honest, reproducible benchmarks.
$8 a month
SelectFund deeper work: Reame performance and roadmap features (warm-ahead prefill, MoE serving) plus TargetBridge compatibility and multi-Mac improvements. Your issues get priority.
$20 a month
SelectBack the big items: ARCA (Reame's shared memory daemon, a "Redis for AI"), first-class MoE serving on cheap CPUs, hardware testing, and long-term maintenance of both Reame and TargetBridge.