Two machines. One principle. The AI you run should sit on your desk, owned by you, in hardware built for it.
32GB of fast VRAM, Qwen 3.8 27B pre installed, and a Quick Start USB to get you up and running. Built for those who want to own their AI usage.
Product image is a render for illustrationCoding, agentic work and repo-level generation, scored against Muse Glimmer and Opus 4.6 Max. Scores are as published by the model vendors.
Green is Qwen 3.8 27B, the model running on FORGE. Muse Glimmer and Opus 4.6 Max are shown for comparison.
Higher is better, scores out of 100. Muse Glimmer has no published score for QwenSWEBench or NL2Repo-Bench.
Private AI infrastructure suitable for enthusiasts or a team. Two GB10 nodes in a purpose built, fan cooled enclosure, assembled and pre configured with DeepSeek V4.1 Flash 1017. Built for those who want serious performance on their network.
Product image is a render for illustration75 average tokens per second on a single stream, 130 across four. Benchmark scores below are as published by the model vendors, against Opus 5.
Green is DeepSeek V4.1 Flash 1017, the model running on FOUNDRY. Opus 5 is shown for comparison.
Higher is better, scores out of 100. V4.1-Flash-1017 leads Opus 5 on three of these four tests.
DeepSeek-V4.1-Flash-1017 Model Card (Hugging Face)
Claude Opus 5 62.7tok/s median based on 16,708 Claude Code turns (AgentsRoom)
Models launch with the press of a button. The appliance sits on your network, so you connect into it from your own machine and work with the tools you already use.
Start, stop and switch models from your computer.
Our appliance is part of your network and accessed through a single endpoint. Point Claude Code, Codex, Hermes or your own tooling at it and work.
We just sell the machine. No telemetry to us, no cloud dependency. Read how we handle data.