I like the idea, but performance is not described. What are the FP8 TOPS? When you pay for a monthly subscription, not only do you performance higher than a local model, but also performance increases as they become available. With this unit (the CPU/RAM specification is also absent) the hardware you get is the hardware you will always have.
If you are going to compare the price of your product directly to the price of a subscription, you are obligated to describe the differences because it is not an apples-to-apples comparison.
You're going to want to include some sort of multi-agent ecosystem because using it for just "chat" or "RAG" is so 2025 ...
Quick context on why I built this: I was personally spending ~$1,200/month on Claude API (I use it for everything — coding, writing, analysis). That’s $14,400/year. Even if I dropped to a single $20/month sub, every query I send tells Anthropic what I’m working on, what I’m thinking about, what problems I have.
I wanted Claude-quality inference without the cloud dependency. That’s not possible yet for large models, but what IS possible locally is surprisingly good.
What’s actually on it:
- 15 chat models — Qwen3-235B-A22B (MoE, fast on consumer hardware), LLaMA 3.3 70B Q4, DeepSeek-R1 32B, Gemma 3 27B, Phi-4
- 14 vision models — drag in an image, get analysis locally. No upload to any server.
- 5 coding assistants — Qwen2.5-Coder 32B is genuinely impressive for local inference
- 3 image generators via ComfyUI — FLUX.1-schnell, Z-Image-Turbo, Qwen-Image. This is the part I haven’t seen anyone else ship as plug-and-play. Most local image gen setups require 30 min of dependency hell.
- RAG/knowledgebase via Qdrant — drop in PDFs, docs, notes; semantic search across them locally
- Medical AI (MedGemma), uncensored models (Dolphin, Nous-Hermes, Abliterated variants)
The target user: Anyone paying $200+/month across Claude, ChatGPT, Midjourney, GitHub Copilot, Perplexity. That’s $2,400/year with zero privacy. The Core tier ($399) pays for itself in 2 months if you cancel just one sub.
Stack: Tauri + Rust + React frontend. Ollama for inference. Qdrant for vector search. ComfyUI for image gen. Everything runs from the SSD — unplug and zero trace on the host machine.
Why not just software? Three reasons:
1. Downloading all of this takes 6-8 hours and 300+ GB — we pre-load
2. It’s portable between machines (home Mac, work PC, travel laptop)
3. Zero install, zero config — double-click and it works
Honest limitations:
- 8GB RAM minimum, 16GB recommended for the larger models
- Not a GPU box — these are quantized models optimized for CPU/unified memory
- Not a replacement for frontier models (GPT-4, Claude Opus) for the hardest tasks
- First batch is 25 units. This is validation.
