localvmAbout this project
localvm-research is an independent research project by Simon-Pierre Boucher (contact@spboucher.ai) investigating whether pretrained large language models that normally do not fit in a consumer Mac's memory can be transformed — strictly post-training — into execution representations with dramatically smaller instantaneous working sets.
The methodology is deliberately strict: an append-only research log, registered hypotheses with explicit falsification criteria before any experiment runs, no results reported from uncommitted code, hardware manifests embedded in every result file, and negative results kept and published. The full methodology is codified in the research charter.
Primary research hardware
| Chip | Apple M5 Max — 6P + 12E CPU cores, 40 GPU cores |
|---|---|
| Unified memory | 48 GB (16 KiB pages) |
| Storage | APPLE SSD AP2048Z 2 TB — measured ceiling ≈ 13.1 GB/s (iostat-validated) |
| OS | macOS 27.0 (26A5388g) |
| Stack | Python 3.14.4, MLX 0.32.0, PyTorch 2.12.0 (MPS), Metal |
What would count as a breakthrough
- A model significantly larger than unified memory running locally with acceptable interactive latency.
- Bytes transferred per token substantially smaller than the compressed checkpoint.
- Evidence that only a small token-dependent fraction of model information is required during typical inference.
- A progressive/conditional execution mechanism that preserves quality while avoiding most weight loading.
- A post-training representation with a qualitatively better storage/RAM/quality tradeoff than fixed quantization.
Failure is an acceptable outcome — the project's charter defines explicit failure criteria, and the log records why an approach died, not just what survived.