Loading video...

Video Failed to Load

Go Home

1/ Today we're introducing Cua-S1-4B-0.2, the first multimodal decision model trained with RLOO on live computer-use tasks, using task-completion rewards. Text and multimodal adapters are available under Apache-2.0:

163,140 views • 1 month ago •via X (Twitter)

39 Comments

Cua's profile picture
Cua1 month ago

2/ At each step, the CUA-S1 model receives the screen state, the task goal, and a fixed set of candidate actions. It returns one action. The environment changes, and the next step starts from the new state.

Cua's profile picture
Cua1 month ago

3/ The training recipe has two stages. Supervised training teaches the decision format. Agentic RL then runs the model in live cua-bench-basic environments, where only a completed task earns the environment's reward.

Cua's profile picture
Cua1 month ago

4/ On the same held-out tasks, Cua-S1-4B-0.2 completes 17/18 text episodes and 13/18 multimodal. Zero-shot djev completes 16/18 and 12/18.

Cua's profile picture
Cua1 month ago

5/ On a frozen GUI-360 split of 168 multimodal tasks, Cua-S1-4B-0.2 reaches 92.9% versus 60.1% for untrained djev. Both use the same inputs and scoring, with no accessibility tree.

Cua's profile picture
Cua1 month ago

6/ The training code and benchmark results are now merged into Cua. Explore the code: Get the Apache-2.0 text and multimodal adapters:

Robert Scoble's profile picture
Robert Scoble1 month ago

What you all are doing is way over my pay grade. So I asked Grok why this is important. Awesome. Its answer: +++++ It matters because it is a concrete step toward computer-use agents that can be trained, inspected, and run outside the largest closed labs. Most computer-use systems still treat the desktop as a chat problem: a large model looks at a screenshot, writes a plan, and hopes the next click works. Cua-S1-4B-0.2 is built as a specialist instead. At each step it gets the screen, the goal, and a fixed list of candidate actions, then returns one action. The environment changes, and the next step starts from the new state. That is closer to how real UI work happens: many small, checkable decisions rather than one long generated essay. The training claim is the other reason people noticed it. Cua says this is the first multimodal decision model trained with RLOO on live computer-use tasks, using task-completion rewards. In other words, the model is not only imitating labeled clicks. After a supervised stage that teaches the decision format, it is rolled out in live cua-bench-basic environments and only scored when the task actually finishes. RLOO is a relatively simple online RL method (REINFORCE with a leave-one-out baseline), so the recipe is easier to reproduce than full PPO-style agent training. The numbers are narrow but not empty. On a frozen 168-task GUI-360 multimodal split, with no accessibility tree, Cua reports 92.9% versus 60.1% for the untrained djev baseline. On held-out live episodes it reports 17/18 text completions and 13/18 multimodal. Those are small, vendor-run evaluations, and Cua itself marks the checkpoint as early research. Pagination is still weak, and at least one live multimodal task is unsolved. So this is not evidence that a 4B adapter replaces Claude or OpenAI computer-use agents. It is evidence that a small, open decision layer can be trained against real GUI outcomes and beat its own untrained baseline by a large margin. That combination is useful in practice. The adapters sit on a frozen Qwen3.5-4B base, ship under Apache-2.0, and are small enough to run locally. Cua’s larger project already supplies the missing pieces around the model: a driver for real desktops, sandboxes, and benchmarks that can generate training trajectories. If the field is going to move from “frontier model remote-controls your screen” to “cheap, specialized policies that operate software,” this is one of the first public artifacts that looks like that stack. The honest limit is transfer. Strong results on Cua’s held-out splits do not automatically mean reliable performance on arbitrary apps, operating systems, languages, or layouts. The importance is the recipe, not a claim that desktop agents are solved: open weights, live-task rewards, one decision per step, and a training loop other people can inspect.

Gognumb's profile picture
Gognumb1 month ago

Love how fast cua team innovate computer use !

Tony Simons's profile picture
Tony Simons1 month ago

Hell yeah!! Congrats on this one Team Cua! 🍾 Fantastically done as always!

Pablo Magana Gabaude 🇫🇷🇨🇱🇸🇪's profile picture
Pablo Magana Gabaude 🇫🇷🇨🇱🇸🇪1 month ago

Will this be a part of Hermes Agent?

Anderson's profile picture
Anderson1 month ago

your mouse just became legacy hardware

Florian S's profile picture
Florian S1 month ago

Amazing work, looking forward to tet your model in (upcoming) Image Jev Bench. Maybe I should start a Computer Use Jev Bench as well

Ziwen's profile picture
Ziwen1 month ago

ain't gonna lie we really do need a computer use model

RabbitHoleExplorer's profile picture
RabbitHoleExplorer1 month ago

re: Cua There is one particularly interesting implication for FirstMate / computer-use agents: the really powerful architecture may not be Jev VS CUA-S1. It could be Jev ABOVE something like CUA-S1. Sol / Qwen / DeepSeek novel planning ↓ Jev task/policy decision ↓ CUA-S1 concrete UI decision ↓ Cua Driver deterministic execution ↓ verification That is starting to look less like “an LLM controlling a computer” and more like an actual inference hierarchy, where the expensive model is invoked only when the lower layers genuinely don't know what to do. @kunchenguid @CompleteSkeptic @typesafeai

JC's profile picture
JC1 month ago

can't wait to give it a run

Panther's profile picture
Panther1 month ago

a 4b model doing computer use means my laptop fan filed for hazard pay

Oliviero Pinotti's profile picture
Oliviero Pinotti1 month ago

🤯 this is crazy powerful, love cua!

Mukhsin Mukhtorov's profile picture
Mukhsin Mukhtorov1 month ago

training on task completion is the part i care about. in my own computer-use loop the model rarely clicked the wrong thing. it said done before the screen had actually changed. curious if completion rewards cut down on that early done

Stephen Brouhard's profile picture
Stephen Brouhard1 month ago

This is a great direction. Not every action needs a giant model thinking through it from scratch. Smaller, purpose built decision models for specific tasks make a ton of sense! 🫡

Anthony Ronning's profile picture
Anthony Ronning1 month ago

Love the specialized CU models and workflows coming out! Curious what the memory profile of this model is for running locally?

Shaurya's profile picture
Shaurya1 month ago

CUA agent focused on task completing is worth testing . Intresting !

Kartik's profile picture
Kartik1 month ago

Cua team is cooking!

pH's profile picture
pH1 month ago

always cooking 🔥

Layton Gott's profile picture
Layton Gott1 month ago

This just get's better and better

Stephen Solka's profile picture
Stephen Solka1 month ago

Stand by. Gpus spinning to add you to leaderboard.

Miguel Saavedra's profile picture
Miguel Saavedra1 month ago

Boom! Let’s goooo

The Real DMT's profile picture
The Real DMT1 month ago

@grok what is the accuracy rate of cua vs jev?

Milind S's profile picture
Milind S1 month ago

This is huge! Fastest computer use ever

Kevin Rajan's profile picture
Kevin Rajan1 month ago

im so hypeeee

Elias Stråvik's profile picture
Elias Stråvik1 month ago

letsgoo super excited to play more with cua - this inspired me to make a launch video with opus 5.5, best one shot i’ve ever seen

Arsh - 16 y/o builder's profile picture
Arsh - 16 y/o builder1 month ago

Every single day you all are shipping something new working at cua must be exhilarating

Kush Agarwal's profile picture
Kush Agarwal1 month ago

Browser and computer use about to be smooth

Brandon Shore's profile picture
Brandon Shore1 month ago

how does this work

AS Maruf 🇨🇦🇧🇩's profile picture
AS Maruf 🇨🇦🇧🇩1 month ago

How do I use Hermes agent to use CUA’s new model? @grok

Brjan | AI Builder's profile picture
Brjan | AI Builder1 month ago

Cua-S1-4B-0.2's focus on live tasks with task-completion rewards is intriguing

peyCyber's profile picture
peyCyber1 month ago

held-out computer-use tasks will reveal whether rloo generalizes beyond training workflows

E_genius's profile picture
E_genius1 month ago

The one-action loop is the interesting bit. Less agent theater, more measurable task completion

Spencer Zhao's profile picture
Spencer Zhao1 month ago

The no-accessibility-tree result is the spicy part. Would love to see a live-site eval where the button moves after the screenshot. That's when browser agents start sweating.

Ben Mo's profile picture
Ben Mo1 month ago

That GUI-360 jump made me look twice. 60.1% to 92.9% without an accessibility tree is a big result. I'm especially interested in what happens after a wrong click. Recovery is where computer agents tend to get messy.

Aden's profile picture
Aden1 month ago

One-episode gains on 18 tasks. GUI-360's 93% vs 60% is the story.

Related Videos

JARVIS-1: Open-World Multi-task Agents with Memory-Augmented Multimodal Language Models paper page: Achieving human-like planning and control with multimodal observations in an open world is a key milestone for more functional generalist agents. Existing approaches can handle certain long-horizon tasks in an open world. However, they still struggle when the number of open-world tasks could potentially be infinite and lack the capability to progressively enhance task completion as game time progresses. We introduce JARVIS-1, an open-world agent that can perceive multimodal input (visual observations and human instructions), generate sophisticated plans, and perform embodied control, all within the popular yet challenging open-world Minecraft universe. Specifically, we develop JARVIS-1 on top of pre-trained multimodal language models, which map visual observations and textual instructions to plans. The plans will be ultimately dispatched to the goal-conditioned controllers. We outfit JARVIS-1 with a multimodal memory, which facilitates planning using both pre-trained knowledge and its actual game survival experiences. In our experiments, JARVIS-1 exhibits nearly perfect performances across over 200 varying tasks from the Minecraft Universe Benchmark, ranging from entry to intermediate levels. JARVIS-1 has achieved a completion rate of 12.5% in the long-horizon diamond pickaxe task. This represents a significant increase up to 5 times compared to previous records. Furthermore, we show that JARVIS-1 is able to self-improve following a life-long learning paradigm thanks to multimodal memory, sparking a more general intelligence and improved autonomy.

AK

141,481 views • 3 years ago