The machine
Hardware
JARVIS runs on one desktop PC. Everything local, from speech recognition to model training, shares this machine.
| Processor | AMD Ryzen 9 9900X, 12 cores Recorded by the full-suite runner for every suite run from August to September 2026. |
|---|---|
| Graphics | NVIDIA GeForce RTX 5070, 12 GB VRAM 12,227 MiB reported. Driver 610.74, CUDA 13.0, BF16 supported (recorded 10 September 2026). |
| Memory | 32 GB system RAM |
| Operating system | Windows 11 Home (build 26200) |
| Local model stack | Python 3.12, PyTorch 2.13 (CUDA 13), Transformers 5.14, PEFT 0.20, bitsandbytes 0.50 An isolated environment, separate from the assistant's own dependencies. |
Why this matters
- Everything shares one 12 GB card: local model training and evaluation, speech recognition, the voice, the HUD and the body services.
- The owner's frozen envelope for local cognition keeps total dedicated GPU use under 11.5 GiB and forbids intentional CPU offload. One heavy GPU job runs at a time; training and inference are serial.
- Latency targets for local cognition (warm first token p95 at most 3 s, at least 15 tokens per second) are acceptance targets, not measured results.
- No serial numbers, host names or account names are published.
As documented 22 September 2026 in docs/training/JARVIS_LM_EXECUTION.md; docs/ledger/SUITE_RESULT.json; docs/ledger/FINAL_LM_PRODUCTION_CAMPAIGN_AUTHORIZATION.md. Specifications come from the project’s own records, not from a fresh inventory.