The Research Fleet
These are the machines the studio runs on. More than anyone should have in a home office. The fleet allows for models up to 230B parameters to be run locally. Many models can be run at maximum context length and BF16. There's headroom for training too: full fine-tunes in the single-digit billions, LoRA and QLoRA adapters up to the 120B class, preference tuning and distillation. And yes, the names come from Tolkien.
[Photograph of the hardware, by Jack. Coming.]
[Second photograph, optional]
Barrowblade
The studio server. Always on and headless; runs the agent harness and dispatches every benchmark attempt.
Radeon 8060S iGPU · Ryzen AI Max+ 395 · 128 GB unified memory
Rohan
Inference node. Serves the local models under study; its two cards pool 128 GB of VRAM.
RTX PRO 6000 Blackwell 96 GB + RTX PRO 4500 32 GB · Ryzen 7 9700X · 64 GB RAM
Andúril
Inference node. Generation and review work, plus the studio's speech recognition.
RTX PRO 4500 Blackwell 32 GB · Ryzen 5 9600 · 32 GB RAM
Annúminas
The workstation. Video editing and daily work. The GPU serves models when a study needs it.
RTX PRO 4500 Blackwell 32 GB · Ryzen 7 7700 · 96 GB RAM
Eregion
The memory host. The studio's corpus and its retrieval live here.
Arc Pro B60 24 GB + Arc Pro B50 16 GB · Ryzen 5 7500F · 96 GB RAM
Imladris
The vault. A ZFS NAS; every machine's nightly backups land here.
Ryzen 5 5600GT · 16 GB RAM · 14.5 TB ZFS pool
Elostirion
The watchtower. A Raspberry Pi with a small rack display; it sees all, or at least the dashboards.
Raspberry Pi 5 · 16 GB RAM
Halbarad
The laptop. Ranges far and returns again.
RTX 4070 Mobile · Ryzen AI 9 HX 370 · 32 GB RAM