M

AutoCAD Bench

Data service

by Markov

Open 50-task benchmark (21 two-dimensional and 29 three-dimensional drafting tasks) in which an AI agent works a sandboxed AutoCAD session through mouse and keyboard to turn a dimensioned reference drawing into an editable DWG file, scored on geometry, dimensions, annotations and presentation; tasks and code are published on Hugging Face.

Find alternatives to AutoCAD Bench

What it does

• Evaluation and observability
• Computer use

Sources

Pricing

Not found · 19 Oct 2026

The page publishes no price or plan; tasks and code are linked on Hugging Face without a charge, and AutoCAD itself needs the user's own licence.

Description

Official documentation · 19 Oct 2026

"A 50-task benchmark for evaluating whether AI models can complete precise AutoCAD work using computer-use alone." "It includes 21 2D and 29 3D tasks." Score: "raw score = 0.60 × geometry + 0.25 × dimensions + 0.10 × annotations + 0.05 × presentation".

What it does

Official documentation · 19 Oct 2026

"AutoCAD Bench tests whether multimodal agents can translate a dimensioned reference drawing into a correct, editable DWG."

Sold within

Official documentation · 19 Oct 2026

Published openly on Hugging Face; no purchase route is involved.

Name

Official documentation · 19 Oct 2026

Page title and heading: "Introducing AutoCAD Bench".

URL

Official documentation · 19 Oct 2026

Markov research page "Introducing AutoCAD Bench", listed on the Research menu beside CAD Environments.

Deployment

Official documentation · 19 Oct 2026

The sandbox is "AutoCAD 2019 on Windows; one clean attempt.dwg is open at handoff." The user runs the harness; the tasks are downloaded from Hugging Face.

Interfaces

Official documentation · 19 Oct 2026

Web research page with Tasks and Code links; the harness is run from the published code.

Modalities

Official documentation · 19 Oct 2026

"Each agent receives a task prompt, a reference image, and a clean AutoCAD document"; "Final DWG, screenshots, model turns, and input events are retained for audit."

Intended for

Official documentation · 19 Oct 2026

A benchmark for model developers and researchers evaluating computer-use agents.

Also from Markov

• Markov computer-use data Data service

  Computer-use training data and environments for AI labs: screen recordings with synced mouse and keyboard events and narrations across browser, coding, design, spreadsheet and CAD workflows, with problem statements, gold output files and rubrics for the CAD environments; open datasets on Hugging Face, and commercial data by contact with no published price.

Open Markov in the directory

Something wrong here? Send a correction — quote this product id: markov-autocad-bench.