NVIDIA Inception member · Accepted July 2026

Every model here is born on NVIDIA.

Type a sentence, get a textured 3D model. That path runs end to end on NVIDIA silicon, and on three.ws the first lane of it is free, keyless, and open source. This page is the honest map of how: which NVIDIA-hosted models do what, which GPUs we run ourselves, and what we learned putting them into production.

Generate one free The model map Engineering post About the membership

Two layers, one stack.

Almost every platform that says "runs on NVIDIA" means one of these. We run both, and they serve different jobs: the hosted lane keeps the free tier free, the self-hosted fleet buys quality the free tier cannot reach.

Layer 1 · Hosted

NVIDIA-hosted models, one key

A single NVIDIA_API_KEY from build.nvidia.com unlocks the entire catalog below. No per-model billing, no per-seat cost. It is rate limited rather than metered, which is exactly why we treat it as a free-first lane and always keep a fallback behind it.

Text to 3D that costs a visitor nothing and needs no account
Chat, vision, embeddings, reranking, safety, speech on the same key
Every call has a documented failover, so a rate limit degrades instead of breaking

Layer 2 · Self-hosted

Our own GPU fleet

Twelve Cloud Run GPU services, each a worker wrapping one model, autoscaled per lane. Eleven on nvidia-l4, one on nvidia-rtx-pro-6000 (Blackwell). This is where the paid realism lanes live, and where PBR materials, rigging, and motion get made.

Physically-based materials: baseColor, metallicRoughness, and normal maps
Auto-rigging and motion, so a generated mesh can move, not just sit there
One wire contract across every worker, so any lane can fail over to any other

The free model map.

Every capability below is in production source and reachable on the free key. Nothing here is aspirational, and nothing is a demo stub.

Capability	NVIDIA model	What it powers	Free
Text to 3D	`microsoft/trellis`	The keyless generate lane on /forge	Yes
Text to image	`black-forest-labs/flux.1-schnell`	Reference images and texture inputs	Yes
LLM, default lane	`meta/llama-3.3-70b-instruct`	Agent chat and platform reasoning	Yes
LLM, model garden	Nemotron 120B / 49B / Nano 9B and peers	The Brain workbench model picker	Yes
Vision	`nvidia/nemotron-nano-12b-v2-vl`	Photo checks before reconstruction, alt text	Yes
Embeddings	`nvidia/nv-embedqa-e5-v5`	Agent memory and semantic search	Yes
Reranking	`nvidia/rerank-qa-mistral-4b`	Sharpening retrieval before an answer	Yes
Content safety	`nvidia/llama-3.1-nemoguard-8b-content-safety`	The guardrail in front of generation	Yes
Text to speech	`magpie-tts-multilingual` (Riva)	Avatar voices, multilingual	Yes

Model by model, with the file each one is wired into: NVIDIA models on three.ws. It names the exact endpoint, the fallback behind it, and the environment variable that turns each lane on.

Try the free lane right now.

No key, no account, no wallet. One POST returns a real GLB you can open in any 3D viewer. This is the same endpoint the browser UI calls.

# generate a textured 3D model on the free NVIDIA lane
curl -s -X POST https://three.ws/api/forge \
  -H 'content-type: application/json' \
  -d '{"prompt":"a small brass compass","backend":"nvidia"}'

# → {"creation_id":"…","status":"done",
#    "glb_url":"https://…/forge/anon/….glb",
#    "quality":{"valid":true,"score":0.947,…}}

The quality block is not decoration. Every generation is inspected before it is handed back: vertex and triangle counts, bounding-box diagonal, whether materials and textures actually landed. A model that fails inspection is regenerated rather than shipped to you broken.

The GPU fleet.

What we run ourselves, and on which card. Each row is a separate Cloud Run service with its own scaling policy, so a slow texture job never queues behind a rigging job.

Lane	Worker	GPU	CPU / memory
Image to 3D, PBR	`model-hunyuan3d-21-rtx`	`nvidia-rtx-pro-6000`	20 / 80 Gi
Image to 3D, fallback	`model-hunyuan3d`	`nvidia-l4`	8 / 32 Gi
Text to 3D	`model-trellis`	`nvidia-l4`	8 / 32 Gi
Image to 3D, fast	`model-triposg`, `model-triposr`	`nvidia-l4`	8 / 32 Gi, 4 / 16 Gi
Texture synthesis	`texture`	`nvidia-l4`	8 / 32 Gi
Auto-rigging	`rig`, `unirig`	`nvidia-l4`	4 / 16 Gi
Text to motion	`model-text2motion`	`nvidia-l4`	4 / 16 Gi
Video to scene	`model-video2scene`	`nvidia-l4`	8 / 32 Gi

Everything above is open source. Each worker's Dockerfile, deploy config, and README are in the repository, including the CUDA and PyTorch versions per GPU generation.

Why this matters.

Generating 3D used to require a workstation, a licence, and a skill you spent years acquiring. The reason a visitor can now type a sentence on a school laptop and get a real, textured, riggable model back for nothing is not one famous model getting good. It is an entire supporting cast getting good and getting free at the same time, most of it hosted by NVIDIA behind a single key. GPU capacity is the one hard constraint on how fast and how detailed that gets, which is exactly the constraint Inception helps relax.

Generate a model View source All partners

three.ws is an independent member of the NVIDIA Inception program, NVIDIA's global program for startups building on accelerated computing. Inception is a startup program, not a partnership, a reseller relationship, or an investment: NVIDIA reviews applicants and admits companies building seriously on accelerated computing. This page describes three.ws products that run on NVIDIA models and hardware; it is not an NVIDIA product and is not endorsed by NVIDIA beyond the program membership shown. NVIDIA, the NVIDIA logo, NVIDIA Inception, NIM, Riva, Nemotron, and NemoGuard are trademarks of NVIDIA Corporation. Learn more about the program at nvidia.com/en-us/startups.

Every model here is born on NVIDIA.

Two layers, one stack.

NVIDIA-hosted models, one key

Our own GPU fleet

The free model map.

Try the free lane right now.

The GPU fleet.

Read next.

Free text-to-3D on NVIDIA

Self-host TRELLIS on your own NIM

Image-to-3D on L4 and Blackwell

The generator was never the hard part

Why this matters.