Harpoon
Log inBook a pilot
Blender expertise for AI labs

Make your model great at Blender.

Blender is where millions of people make 3D: games, films, products, the scenes robots train in. Harpoon's Blender professionals test your model, correct it and give it the data it needs to work in Blender like an expert, and make extraordinary things.

Book a pilotJoin as a Blender expert
See the kind of tasks we test
Object Mode
chair.blend
User Perspective
(1) Chair | Seat
4 defects found
Example · chair from a model's Blender script
  1. 1Front-right leg never built: the loop stops at three.
  2. 2Lower back slat sits inside the seat.
  3. 3Top slat sticks out through both posts.
  4. 4Wrong units: in Blender the chair is 9 m tall.
Drag to turn the chair.
Built forCoding modelsComputer-use agentsText-to-3D3D generationWorld modelsRobot simulation

Ask a model for a Blender script and it writes one in seconds. Run it, and the chair is missing a leg, the scene is ten times too big and the arm tears at the elbow. The code looks fine. Only Blender, and someone who uses it every day, can tell.

All of Blender

From the first cube to the final render.

Each part of Blender takes its own skill. Every task goes to an expert who does that work for a living.

ModelingShapes, real-world size, clean topology, parts that touch
UVs and materialsUnwrapping, node trees, textures that line up
RiggingBones, weights, joints that bend without tearing
AnimationKeyframes, timing, curves, motion that reads
Geometry NodesProcedural setups that hold up when inputs change
Lighting and renderingCycles and EEVEE, cameras, light, render settings
Python scriptingScripts and add-ons that run and build the right thing
Using the interfaceAgents that click through Blender like an artist
What labs get

Three ways we make your model better at Blender.

01RLHF

Expert evaluation

Your model writes a Blender script, drives Blender as an agent or makes an asset. We run it in Blender, then a professional scores it, compares it, marks every defect and fixes what matters.

  • Scores with written reasons
  • Preference pairs (A or B, and why)
  • Located defects and corrected files
02Tasks

Blender tasks and environments

Real work, written by people who do it: from a simple table to a full character rig. Each task comes with an ideal solution and a check that runs in Blender and scores the result.

  • Tasks from easy to expert, with solutions
  • Automatic checks for reinforcement learning
  • Private test sets your model has never seen
03Data

Training data

Recorded sessions of experts working in Blender, with every action and the reason behind it, and clean, annotated .blend files your model can learn from.

  • Expert sessions: screen, actions, reasoning
  • Annotated .blend and asset datasets
  • Before and after files for every fix

Every output is run in a real copy of Blender before anyone looks at it. How it works for labs

How it works

Blender checks first. Then an expert.

  1. 01

    Run in Blender

    Every script, file or agent session runs in a clean copy of Blender. We measure what it really built, in seconds.

  2. 02

    Expert review

    A Blender professional looks at what a program can't judge, scores it and says why.

  3. 03

    Fix and deliver

    The outputs that matter are corrected. You get scores, comparisons and clean files.

Every expert passes a practical test in Blender before they start and is scored on hidden test questions while they work.

Task 0412 · Blender scripting · Compare06:41

Prompt: “Write a Blender Python script for a wooden chair: seat at 45 cm, four legs, two back slats.”

A
seat = add_box(4.5, 4.5, 0.4)
seat.location.z = 4.3
for x, y in legs[:3]:
    add_leg(x, y)
Runs · 3 of 4 legs · 9 m tall
BPreferred
seat = add_box(0.45, 0.45, 0.04)
seat.location.z = 0.43
for x, y in legs:
    add_leg(x, y)
Runs · 4 legs · 0.9 m tall
CriterionAB
Script runs
Matches the request
Clean, usable mesh
Why B is better
Both scripts run, but A builds only three legs and mixes up its units, so the chair comes out ten times too big. B is the right size. Its slats are a little thin for wood.
Annotations

Every review becomes training data.

An expert doesn't just say “bad”. Each review is a structured record your team can train on directly.

Scores on a rubric
Each output is rated on what matters in Blender: does it run, does it match the request, is the result usable.
Defects, located
Every problem gets a type, the exact object or line of code, and how to fix it.
Preferences with reasons
Which of two outputs is better, and a written reason a model can learn from.
Corrected files
The expert's fixed script or .blend, so you have the before and the after.
Run records
Blender version, time to run, errors and what the scene contains after the run.
annotation_0412.jsonexample
{
task: "0412",
area: "blender scripting",
output: "chair_A.py",
run: { blender: "4.5 LTS", ok: true, seconds: 1.8 },
scores: { runs: 5, matches_request: 2, clean_mesh: 3 },
defects: [
{ type: "missing_part", where: "front-right leg", line: 3 },
{ type: "wrong_units", where: "seat, 4.5 m wide", line: 1 },
{ type: "intersecting_mesh", where: "lower slat / seat" }
],
preferred: "B",
reason: "A builds three legs, ten times too big.",
fixed_file: "chair_A_fixed.py",
reviewer: { skill: "blender modeling", test_accuracy: 0.96 },
}

Example record from the review above. Delivered as JSON, with the fixed files.

Harpoon Blender Bench

Real Blender work, graded by Blender.

Tasks written by professionals, from level 1 to level 3. The model's answer runs in Blender and a check scores what it built, so results can't be faked by code that only looks right.

Put your model on the bench
L1ModelingA dining table, 120 × 80 cm, 75 cm tall, four legs.Size, part count, legs touch the top, scale applied
L1ScriptingAdd 12 cubes in a circle, each turned to face the center.Script runs, positions, rotations
L2MaterialsGive the mug a glossy red ceramic material and clean UVs.Node setup, UV overlap and stretch
L2Geometry NodesScatter 200 rocks on the terrain. None may float.Count, contact with the ground
L3RiggingRig this arm with three bones. The elbow bends 90° without tearing.Bones, weights, mesh at 90°
L3AnimationA ball bounces three times in two seconds, each bounce lower.Keyframes, heights, timing at 24 fps
L3AgentOpen the scene, fix the flipped normals, save the file.Normals and contents of the saved file
Free tool

Check a 3D model for free.

Drop a .glb exported from Blender or made by any text-to-3D tool. You get a score in seconds. No account, and the file never leaves your computer.

Drop a .glb file here
or click to choose one · stays on your computer
For AI labs

Start with a small pilot.

Send us a batch of what your model makes in Blender. You get back scores, expert reviews and fixes, and you decide if it's worth going further.

Book a pilot
For Blender experts

Get paid to teach AI Blender.

Modelers, riggers, animators, technical artists. Remote, on your own hours, paid by the hour.

Apply as an expert
Our mission

Make AI a true Blender professional, then bring the same expertise to every tool people use to build in 3D.

The Blender benchmark everyone cites

A public, expert-built measure of how well AI models use Blender, area by area.

The first source of Blender training data

Evaluations, tasks and expert sessions that AI labs trust and train on.

The largest network of Blender experts

Modelers, riggers, animators and technical artists, tested, scored and well paid.

The road ahead
  1. Now

    Evaluate and train on Blender

    • Expert evaluation of Blender scripts and assets
    • Blender tasks with automatic checks
    • Annotated .blend and asset datasets
  2. Next

    Agents and the benchmark

    • The Harpoon Blender Bench: a public ranking of AI models
    • Recorded expert sessions for computer-use agents
    • Original scenes and assets made by our experts
  3. Later

    Every tool that builds in 3D

    • Houdini, Maya, Unreal and other creative tools
    • CAD for engineering and manufacturing
    • Robotics, simulation and the physical world