TURN AN IDEA INTO A PRINTABLE 3D DESIGN.

Describe the part. BullPrintLab infers the dimensions, clearances and manufacturing assumptions, builds real parametric geometry, and lets you change any of it by typing. Every design is a spec, not a picture.

DIMENSIONS AUTO MATERIAL PETG METHOD FDM TOLERANCE 0.4 MM

ADVANCED FIELDS ARE OPTIONAL — EVERY ASSUMPTION IS SHOWN AND EDITABLE AFTER GENERATION

TRY ONE

{ "name": "Wall Bracket", "units": "mm", "manufacturing_method": "FDM", "material": "PETG", "parameters": { "width": 100, "height": 50, "thickness": 4, "wall_thickness": 3, "clearance": 0.4 }, "features": [ { "type": "base_plate", "width": 100, "height": 50, "thickness": 4 }, { "type": "hole_pattern", "hole_type": "through", "diameter": 5, "count": 4, "spacing": 70 } ] }

SOVEREIGN COMPUTE · GPU-01 + GPU-02 · ON OUR FLOOR

ONE CARD HOLDS THE WHOLE MODEL.

BrAhMa runs on Muse-Glimmer-30B, resident on the 96 GB card at Q8_0 and never leaving it — 28 GB of weights against 96 GB of board, so there is no eviction, no cold load between jobs, and 66 GB still free for context and a second model. A 24 GB RTX 3090 registers as gpu-02 and takes overflow at a lower quantisation, so a queue degrades precision rather than dropping the job. Geometry itself never touches the GPU: OpenSCAD and CadQuery run deterministic on the CPU, so a render burst can't block a part from generating.

THE RACK

GPU-01RTX PRO 6000 BLACKWELL · 96 GB · BrAhMa RESIDENT

GPU-02RTX 3090 · 24 GB · OVERFLOW

CPUXEON W9-3475X 36C/72T · RYZEN 9 5900X 12C/24T

RAM251 GB + 62 GB DDR5

STORAGE1.8 TB + 915 GB NVME · MODELS + ASSETS

RUNTIMELLAMA.CPP B2271 · CUDA 12.8 · SM_120

QUEUEREDIS · CELERY · CAPABILITY-ROUTED

ISOLATIONCAD RUNS IN A NO-NETWORK CONTAINER

WORKER REGISTRATION — SCHEDULER ROUTES BY CAPABILITY

{
  "worker": "gpu-01",
  "gpu": "RTX PRO 6000 Blackwell",
  "vram": 96,
  "cpu": "Xeon w9-3475X",
  "ram_gb": 251,
  "capabilities": [
    "text",
    "vision",
    "structured-output"
  ]
}

Adding gpu-02 is a registration, not a rewrite — the app never couples to one card. Generated code executes in a container with no network, hard time and memory limits, and file-size caps.

VRAM BUDGET — GPU-01 · 96 GB, FULLY RESIDENT Q8_0 · llama.cpp

Muse-Glimmer-30B · Q8_0

INTERPRETER — the model under BrAhMa: prompt → BullSpec → OpenSCAD

28 GB

KV CACHE · 16K CTX

Long design conversations without re-reading the spec each turn

2 GB

HEADROOM

Context growth, a perception encoder, and a second model without eviction

66 GB

THE MODEL CALL

Muse-Glimmer-30B at Q8_0 is the model under BrAhMa — near-lossless, because BullSpec generation is a structured-output problem where precision on the numbers matters more than parameter count. It is too new for vLLM and for Ollama's bundled runtime, so it runs on llama.cpp built from master against CUDA 12.8 for sm_120; that is the price of being early, and it is paid once. Qwen3.8-27B holds gpu-02 at Q4_K_M so overflow degrades quantisation rather than dropping the job. Concept render and image-to-3D are specified but not deployed — no card in the fleet holds them today. No open model emits trustworthy triangles — the models write specs, the CAD engine writes geometry. That split is the product.

JOB PIPELINE QUEUED → INTERPRETING → GENERATING_GEOMETRY → RENDERING → VALIDATING → COMPLETE → STREAMED TO THE UI OVER SSE

TURN AN IDEA INTO A PRINTABLE 3D DESIGN.

PROJECTS

ONE CARD HOLDS THE WHOLE MODEL.

TURN AN IDEA INTO A PRINTABLE 3D DESIGN.

PROJECTS

ONE CARD HOLDS THE WHOLE MODEL.