Current model: CM-7, with state-conditioned input and a separate iterative seven-ply live helper. Current evaluation · Current source and reproduction guide. Earlier reports and full-history packages below keep their historical provenance.
Local research hardware
The workstation behind LittlePly.
Verified local configuration, recorded 28 October 2026.
LittlePly is an experiment in AI research guided by Andrei Savu, working with his OpenAI dot, Glimmer, on a local Lambda Vector Workstation.
“Lambda Vector Workstation” is the owner-supplied designation. The specifications below come from local operating-system and NVIDIA inventory, cross-checked against the existing training records; they are not inferred from a sales listing.
| CPU | AMD Ryzen Threadripper PRO 5955WX — 16 cores, 32 logical processors |
|---|---|
| GPUs | 2 × NVIDIA GeForce RTX 4090; 24 GB VRAM per GPU (24,564 MiB reported by NVIDIA per device) |
| System RAM | 128 GiB installed; about 127.8 GiB reported usable by Windows |
| Linux-visible RAM | About 62.7 GiB at inventory time; WSL-visible memory differs from installed host RAM |
| Host OS | Windows 11 Pro |
| Research environment | Ubuntu 22.04.5 LTS under WSL2 |
| Verified Python environment | Python 3.12.15; PyTorch 2.9.0+cu128; PyTorch CUDA runtime 12.8; Transformers 5.21.0 |
Compute boundaries
The recorded full-month training run used both GPUs. GPU memory remains separate per device; this is not a single 48 GB memory pool. Individual experiment records describe their actual precision, timing and configuration.
The earlier GPU serving snapshot uses one RTX 4090 (device index 1 on this workstation), shared across up to four admitted games; bullet games run alone. Training and serving can contend for GPU memory and compute, so training must reserve serving headroom. The earlier CPU-only service remains available as historical source. Serving depends on the local Windows/Ubuntu host remaining awake and running; no continuous-availability guarantee is implied.
These are inventory and recorded configuration facts, not a throughput promise or a new benchmark. No serial numbers, host identifiers, private paths or network configuration are published.