Blender-Bench benchmarks LLMs — and the harnesses that drive them — at using Blender through the Blender MCP tool interface. Each contestant gets a task and a real Blender, and has to build the scene entirely through tool calls: no human hands on the mouse.
Runs are scored by pairwise human voting. You compare two renders of the same task and pick the better one; a Bradley–Terry model turns those votes into the ranking on the leaderboard.
The project was inspired by MineBench, which has always fascinated the author as someone who has worked in 3D for quite some time. We've reached the point where models are advanced enough to finally do meaningful work in 3D.
Understanding 3D is an important capability for AI, especially on the way to world models, AGI, and embodied AI. It comes down to recognizing relationships between objects in space, predicting how they would interact with one another, and developing a sense of visual beauty and composition.
The first wave of Blender-Bench is creative tasks — detailed prompts describing a scene. A second wave will target real-world Blender workflows and everyday 3D work: rigging, retopology, animation. Some parts, like retopology, may be a little overkill — but they're an interesting way to observe how models approach problem-solving. This is only the beginning.
Every contestant works from an authored prompt. Browse them in the task library — each task shows its prompt in a Blender Text Editor and lists the runs recorded against it. Browse results in the gallery or read individual transcripts in the run log.
More is coming: a Second Wave task suite for real production skills (retopology, UV texturing, rigging, rig animation, scene cleanup); 3D Bench, a sibling benchmark for three.js and p5.js scene generation; and BYOK — bring-your-own-key runs so you can put a model through the bench yourself. All still in the works; see the roadmap on the home page.
Built on Blender + Blender MCP. Source on GitHub — and if the bench is useful to you, you can support it. The leaderboard, gallery, and run log read live benchmark runs and votes; the landing battle demo is still illustrative.