SkyForge · hardware pilot
Three desk-side machines, each with enough unified memory to run frontier open-weights models locally — DeepSeek V4 Flash and Pro, Kimi K3, Llama 4 Scout, Qwen. Tell us which box and which models you'd run, and we'll reach out as pilot allocations open.
Mac Studio M5 Ultra · DGX Spark · DGX Station · 8 catalog models
01The machines
Each card lists the catalog models that fit in the machine's memory pool, with the sizing rationale. The same picker drives the form below.
128GB unified memory
GB10 Grace Blackwell on the desk — CUDA on 128GB of unified memory.
Pilot allocation, first-come
Qwen3.8 27B
27B dense at bf16 (~54GB) fits with headroom for a 256K-token KV cache.
DeepSeek-R1-Distill 32B
32B dense at bf16 (~64GB) — step-by-step reasoning on a single unit.
Gemma 3 27B
27B dense at bf16 (~54GB) runs on the 128GB pool with room to spare.
748GB coherent memory
748GB of coherent memory — the desk-side shape for the biggest models.
Limited pilot allocation
DeepSeek V4 Pro
1.6T MoE with a 1M-token context — nothing smaller than this pool holds it.
Kimi K3
2.8T MoE at 4-bit fits within 748GB for agentic workloads.
Qwen3.5 397B-A17B
397B MoE at FP8 (~400GB) fits with headroom for long-context KV cache.
512GB unified memory · 1.2TB/s
Apple silicon with a 512GB unified pool — the personal frontier box.
Announced Aug 25, 2026 · 512GB configs ship ~late Oct
DeepSeek V4 Flash
284B MoE at Q8 (~162GB) fits with room for long-context KV cache.
Llama 4 Scout
109B MoE (17B active) at FP8 runs comfortably inside the 512GB pool.
Qwen3.8 27B
27B dense at bf16 (~54GB) leaves most of the pool for batch and context.
02Join the waitlist
Allocations are first-come, sized by the hardware and models you pick. No spam — just a note when your spot opens.