Rohan Paul (@rohanpaul_ai) on X

What it claims

Hi3D has open-sourced Hi3D v1.0, an image-to-3D model formerly called Sparc3D. According to the thread, the model already ran in Hi3D's live product before its open release, so the open weights are a model that has been used in production rather than a research preview. Rohan Paul describes it as "last year's cutting-edge technology, open-sourced this year." Hi3D's own closed model is at v3.0 and refines at 2048 cubes per side. The open v1.0 is the base that the closed versions build on.

Hi3D also says it has cut users' credit costs on its hosted service by over 50%. This figure comes from Hi3D, and the thread gives no before or after prices.

Evidence

Thirteen people with professional 3D-algorithm experience compared Hi3D v1.0 side by side against eight other image-to-3D models. Across 562 judgments, they picked Hi3D v1.0 83.5% of the time. The thread names only one of the eight competitors, TRELLIS.2.

The team graded the back and bottom of every result, not just the view shown in the input image. The thread gives one example: on a winged armored figure, TRELLIS.2 collapsed into a flat panel, while Hi3D v1.0 kept the wings and limbs as connected parts.

How to use it

Hi3D offers two hosted ways to use its models beyond the open weights:

• The Hi3D Agent combines generating, editing and processing 3D models in one flow.
• A Model Context Protocol (MCP) server lets other AI agents call Hi3D directly.

The thread does not say which model version either of these serves.

Limits

The thread is promotional and thin on detail:

• It gives no weights location, licence, model size, hardware requirements or run commands.
• It does not explain the architecture or what v3.0 changes over v1.0.
• It names only one of the eight baseline models.
• It does not say how the 562 judgments were split between those models or how the evaluators were chosen.
• The single TRELLIS.2 comparison is an anecdote, not a measured rate of failure.