RL Environments Enterprise Huzzle World Talk to founders opens Calendly

Huzzle World: The Frontier of Small Computer-Use Models.

HuzzleWorld is Huzzle Labs’ family of computer-use models, released at 1B, 2B, 4B and 8B parameters and built on our proprietary architecture, our own RL environments, and training from video. On ScreenSpot-Pro, they are state of the art for their size. HuzzleWorld-2B outscores grounding models up to 36 times its size.

Benchmarks 01 / 04

#1 on ScreenSpot-Pro among models of 2B and smaller.

HuzzleWorld-2B Competing 2B models
20 30 40 50 60 70 HuzzleWorld-2B (Zoom In) · 64.2% 64.2 HuzzleWorld-2B (Zoom In) MAI-UI-2B (Zoom In) · 62.8% 62.8 MAI-UI-2B (Zoom In) HuzzleWorld-2B (single pass) · 60.8% 60.8 HuzzleWorld-2B (single pass) UI-Venus-1.5-2B · 57.7% 57.7 UI-Venus-1.5-2B MAI-UI-2B · 57.4% 57.4 MAI-UI-2B GUI-Actor-2VL-2B · 36.7% 36.7 GUI-Actor-2VL-2B UI-TARS-2B · 27.7% 27.7 UI-TARS-2B UGround-V1-2B · 26.6% 26.6 UGround-V1-2B
All 8 leaderboard entries at ≤2BScreenSpot-Pro accuracy %

#1 on OSWorld among models of 8B and smaller.

ScreenSpot-Pro measures one click on a static screen. OSWorld Verified measures whether the whole task actually gets done: a real Ubuntu desktop, real applications, and a script that checks the end state. HuzzleWorld-8B completes 58.6% of them, ahead of every model of 8B or smaller on the leaderboard, and ahead of GUI-Owl-1.5 32B (55.4%), DeepMiner-Mano-72B (53.9%) and opencua-72b-preview (46.1%).

#18BHuzzleWorld-8B58.6%
#28BEvoCUA-8B-2026010546.1%
#37BDART-GUI-7B-092440.5%
#47BDeepMiner-Mano-7B40.1%
#57Bgui-owl-7b32.1%
#67BTianXi-Action-7B30.4%
#77Buitars-1.5-7b29.6%
#87Bopencua-7b28.8%
#97Bopencua-qwen2-7b23.1%

Every model on the OSWorld Verified leaderboard with a published parameter count of 8B or smaller, best run per model. Bars are relative to HuzzleWorld-8B. Across the whole leaderboard, regardless of size, it stands #11 of 41 models.

HuzzleWorld has moved the frontier for small CUA models.

Every leaderboard entry with a published parameter count, plotted from the largest models on the left to the smallest on the right. The line is the best score achievable at each size. At 2B, HuzzleWorld sits on it.

Leaderboard entry HuzzleWorld-2B Best score at this size or smaller
20 30 40 50 60 70 80 400B 200B 100B 50B 20B 10B 5B 2B AdaZoom-GUI-4B (cond. zoom-in, refine w/ Qwen3.5-397B) · 76.8% · 401B Holo2-235B-A22B (Agentic) · 78.5% · 235B Holo2-235B-A22B · 70.6% · 235B Holo1.5-72B · 63.3% · 72B UI-Venus-72B · 61.9% · 72B GTA1-Qwen2.5VL-72B · 58.4% · 72B Qwen2.5-VL-72B-Instruct · 53.3% · 72B UGround-v1-72B · 34.5% · 72B UI-TARS-72B · 38.1% · 72B Duvo Eye-1 (Holo-3.1-35B-A3B + LoRA) · 72.9% · 35B KV-Ground-8B + Qwen3.5-27B Consistency Router · 80.9% · 35B MAI-UI-32B (MVP) · 77.5% · 32B MAI-UI-32B · 67.9% · 32B MAI-UI-32B (Zoom In) · 73.5% · 32B MVP_Qwen3VL-32B · 74.1% · 32B GTA1-32B · 63.6% · 32B GTA1-Qwen2.5VL-32B · 53.6% · 32B Qwen2.5-VL-32B-Instruct · 48.0% · 32B UI-Venus-1-5-30B-A3B · 69.6% · 30B Holo2-30B-A3B (Agentic) · 75.2% · 30B Holo2-30B-A3B · 66.1% · 30B KV-Ground-GuiOwl1.5-0315-8B-ZoomIn · 80.5% · 8B KV-Ground-GuiOwl1.5-0315-8B · 73.2% · 8B UI-Venus-1-5-8B · 68.4% · 8B Holo2-8B (Agentic) · 71.4% · 8B MAI-UI-8B (Zoom In) · 71.9% · 8B MAI-UI-8B · 65.7% · 8B MVP_Qwen3VL-8B · 65.0% · 8B Holo2-8B · 58.9% · 8B BAMI-7B · 57.7% · 7B UI-AGILE-7B · 47.9% · 7B GTA1-7B · 55.5% · 7B GUI-ARP-7B · 60.8% · 7B Holo1.5-7B · 57.9% · 7B V2P-7B · 52.5% · 7B UI-Venus-7B · 50.8% · 7B TianXi-Action-7B · 51.9% · 7B GUI-Actor-2.5VL-7B · 44.6% · 7B GUI-Actor-2VL-7B · 40.7% · 7B GTA1-Qwen2.5VL-7B · 50.1% · 7B SE-GUI-7B · 47.2% · 7B Qwen2.5-VL-7B-Instruct · 26.8% · 7B UI-TARS-7B · 35.7% · 7B Aguvis-7B · 22.9% · 7B UGround-V1-7B · 31.1% · 7B AdaZoom-GUI-4B (conditional zoom-in) · 70.6% · 4B AdaZoom-GUI-4B · 61.6% · 4B KV-Ground-Qwen3VL-4B-ZoomIn · 70.3% · 4B KV-Ground-GuiOwl1.5-4B-0228-ZoomIn · 76.4% · 4B KV-Ground-Qwen3VL-4B · 63.2% · 4B KV-Ground-GuiOwl1.5-0228-4B · 67.0% · 4B Holo2-4B (Agentic) · 68.6% · 4B Holo2-4B · 57.2% · 4B GUI-AIMA-3B · 59.6% · 3B UI-AGILE-3B · 45.0% · 3B Holo1.5-3B · 51.5% · 3B Di-GUI-3B · 31.6% · 3B ZonUI-3B · 28.7% · 3B GUI-Actor-2.5VL-3B · 42.2% · 3B SE-GUI-3B · 35.9% · 3B UI-Venus-1-5-2B · 57.7% · 2B MAI-UI-2B (Zoom In) · 62.8% · 2B MAI-UI-2B · 57.4% · 2B GUI-Actor-2VL-2B · 36.7% · 2B UI-TARS-2B · 27.7% · 2B UGround-V1-2B · 26.6% · 2B HuzzleWorld-2B · 60.8% HuzzleWorld-2B (Zoom In) · 64.2% HuzzleWorld-2B
Parameters, large to small · log scaleScreenSpot-Pro accuracy %

Moving right means a smaller model. The frontier steps down as size falls, and HuzzleWorld-2B is the point where it ends: nothing smaller scores higher.

Leading computer use in enterprise software.

HuzzleWorld-2B against two 72B grounding models across all six ScreenSpot-Pro domains. It leads in every one, at 1/36th the parameter count.

HuzzleWorld-2B Qwen2.5-VL-72B-Instruct UI-TARS-72B
OfficeWord · Excel · PowerPoint
74.8
72.6
54.8
DevelopmentVSCode · PyCharm · Android Studio
63.9
53.5
40.8
ScientificMATLAB · Origin · Stata
61.8
59.1
45.7
Operating systemsWindows · macOS · Linux
61.7
49.5
30.1
CreativePhotoshop · Blender · Premiere
54.8
44.9
39.6
CADAutoCAD · SolidWorks · Vivado
51.0
44.4
17.2

HuzzleWorld-2B against two 72B grounding models across all six ScreenSpot-Pro domains. It leads in every one, at 1/36th the parameter count. Bars are percentage accuracy; values are the published single-pass leaderboard results.

Demos

Watch it work.

HuzzleWorld-2B can take screenshots, return coordinates, click, scroll and type.

File manager 15 stepsCreate two folders and inspect one
Calculator 12 stepsCompute ((7 × 8) − 16) ÷ 4
Scientific calculator 12 stepsChain √256 × 3 + 2, negated
Text editor 17 stepsSave under a new name, then find and replace
Spreadsheet 24 stepsEnter quarterly figures and total them
In production

Train on your recordings. Run on your infrastructure.

The same properties that make HuzzleWorld small make it practical to deploy inside a company rather than behind someone else’s API.

Your data

Your recordings are the training set

Because the models train from video, we can train directly on the explainer recordings your team already has, and deliver working agents for those workflows. No annotation project first.

Your budget

A fraction of the data and compute

A small model needs far less data, compute and time to train than a large one, while matching its accuracy on the workflows you care about. Shorter path from recording to running agent.

Your infrastructure

Nothing leaves the organization

1B to 8B fits on hardware enterprises already run. Deploy HuzzleWorld inside your own environment, so screen recordings and workflow data never leave it.

→ How we deploy this Enterprise
Questions

HuzzleWorld, in short.

What is HuzzleWorld?

HuzzleWorld (also written Huzzle World) is Huzzle Labs’ family of computer-use models, released at 1B, 2B, 4B and 8B parameters. Each takes a screenshot and an instruction and returns the exact screen coordinate to act on, which is the grounding step underneath any computer-use agent.

How does HuzzleWorld-2B score on ScreenSpot-Pro?

60.8% single-pass and 64.2% with an optional two-pass zoom-in. It reaches 74.8% on office software and 70.3% on text-labelled targets. Both settings are the highest recorded for a 2B model on the official leaderboard.

Which larger models does it outscore?

Single-pass, HuzzleWorld-2B outscores grounding models up to 36 times its size, including UI-TARS-72B (38.1%), UGround-v1-72B (34.5%), Qwen2.5-VL-72B-Instruct (53.3%), and H Company’s Holo2-8B (58.9%) and Holo1.5-7B (57.9%).

What is ScreenSpot-Pro?

A GUI grounding benchmark built from full-resolution screenshots, up to 3840 × 1080, of 26 real professional applications. A prediction only counts if it lands inside the ground-truth box. There is no partial credit.

Why does model size matter for computer use?

Grounding runs on every step of an agent trajectory, so its cost and latency multiply across a workflow. A 2B model that matches or beats a 72B model changes where computer-use agents can be deployed and what they cost to run.

How is HuzzleWorld trained?

On our proprietary architecture, in our own long-horizon RL environments, and from video. Learning from screen recordings instead of large hand-annotated datasets means the models can be trained directly on a company’s existing explainer recordings.

Can we run it on our own infrastructure?

Yes. At 1B to 8B the models fit on hardware enterprises already run, so HuzzleWorld can be deployed inside your own environment, with no data leaving the organization.

What is OSWorld Verified?

A benchmark that runs an agent on a real Ubuntu desktop across applications like Chrome, GIMP, LibreOffice, VLC and VS Code, then runs a script to check whether the task was actually completed. Unlike ScreenSpot-Pro it scores end-to-end task completion, not a single click.

How does HuzzleWorld-8B score on OSWorld Verified?

It completes 58.6% of tasks, the highest of any model on the leaderboard published at 8B parameters or smaller, and ahead of GUI-Owl-1.5 32B, DeepMiner-Mano-72B and opencua-72b-preview. Across the leaderboard regardless of size it stands 11th of 41 models.

What is GUI grounding?

The step that turns an instruction and a screenshot into a screen coordinate: deciding exactly where to click. It is the part of a computer-use agent that connects a plan to an action, and it runs on every step of a workflow.

Work with us

Deploying computer-use agents, or training your own?
Let’s talk.

Talk to founders → RL Environments For AI labs → Enterprise Custom evals & models

Leaderboard data as of 30 July 2026. Ranks are positions among the 72 ScreenSpot-Pro leaderboard entries with a published parameter count and a score of 10 or above. Charts are plotted from 20% upwards; four entries scoring below 20% fall outside the plotted range. Scores are grounding accuracy across the benchmark’s 23 applications and three operating systems; the benchmark measures localisation on a static screen, not end-to-end task completion. Single-pass and zoom-in are recorded as separate leaderboard entries. Figures for HuzzleWorld-1B and HuzzleWorld-4B are our own measurements on the same benchmark, with leaderboard submissions in review. OSWorld Verified figures refresh automatically from the public OSWorld leaderboard and show the best recorded run per model; HuzzleWorld-8B’s own score is Huzzle-reported. OSWorld publishes no parameter count, so the size-class board lists only models whose published name states a size of 8B or smaller, and models of undisclosed size are left out of it.