• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    Yutori Releases Navigator N2 Computer-Use Model

    Model scores 65.2 percent on OSWorld 2.0 at low cost and leads multiple agent benchmarks.

    DP
    TA
    AD
    13 Sources, 35d ago, first seen 35d ago

    TLDR

    Yutori team members including Rui Wang and Devi Parikh posted on X that Navigator n2 is now available. Yutori's announcement posts describe the model as selecting among GUI, CLI, tools, and code interfaces to complete long-horizon desktop tasks. Company posts report 65.2 percent on OSWorld 2.0 at $1.46 per task, 85.3 percent on OSWorld-Verified, 82.6 percent on MyPCBench, and 83.1 percent on MacAgentBench. Additional posts note unplanned state-of-the-art results and describe the system as frontier-level across computer-use agent evaluations.

    Combined views

    42.4K

    13 Sources, first seen 35d ago

    Combined views

    42.4K

    13 Sources, first seen 35d ago

    348 likes
    348 likes
    23 comments
    199 saves
    52 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    Featured Source
    23 comments
    199 saves
    52 reposts
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    Navigator n2

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    13 Sources

    @theruiwangFrontier-level computer use: 65.2% on OSWorld 2.0 at $1.46 per task
    @varibandiNavigator N2 should be the default CUA model you try, but dont listen to me listen to the benchmarks! It also costs an order of magnitude less than Opus :) https://yutori.com/blog/introducing-n2
    @_krunalshah_We're launching Navigator n2, a computer-use model that connects with computers deeper than just graphical interface. We've put in a lot of ❤️ and 🤔 into making the model particularly shine on long-horizon computer-use tasks, go check it out at https://yutori.com/n2!
    @abhshkdzWhen we started Yutori, our research goal was to post-train the best browser-use models for powering web agentic products. We did that with Navigator n1 and n1.5; models that were pareto-dominant (at the time of release) in accuracy and cost for browser-based workflows. And browser-use models are basically computer-use models (screenshots in, actions out) optimized for driving web browsers (only), i.e. the data and action sets are constrained to browser control specifically. With today's release of Navigator n2, we've broadened our ambition and gone beyond the browser, towards a computer-use model that can operate full-desktop environments, interleaving graphical (including browsers!) and command-line interfaces, and even writing short snippets of code to complete tasks. And still pushing the accuracy-cost pareto, this time for computer use. General-purpose frontier models ofc have a place in the world, particularly for discovering use-cases that weren't possible say 6 months ago, but if you're looking to ship computer-use / knowledge work agents today at scale, n2 is a compelling option at a fraction of the cost. We spoke with @alexrkonrad to share the thinking behind n2.
    @skanda_vaidNavigator n2 is out! n2 is trained to use computers the "computer way" and is frontier level across all CUA benchmarks! A lot of cool learnings along the way that we're hoping to share soon!
    @svpinoThe issue I have with computer-use models today is that they spend too much time clicking through interfaces. You can die of boredom waiting for these tools to finish a simple task. n2 is a new 27B-parameter computer-use model that handles this differently. When solving a problem, the model chooses how to complete each step based on the fastest available option. The model works in 4 different ways: 1. Use a GUI to interact with an application 2. Write terminal commands to work with the operating system 3. Call an API to access a service directly 4. Write code to handle a task programmatically As n2 works through a task, it switches between these options, and the end result is a much faster solution! For example, to complete a specific workflow, the model may browse the web, open an application, edit one file, run a few commands, and call an API. This is key when you are building agents that work across different tools. You want these agents to be faster, less fragile, and be able to work for longer. You can test the API and interactive playground at http://yutori.com/n2. Thanks to the Yutori team for partnering with me on this post.
    @deviparikhRT @abhshkdz: Excited to release Navigator n2, now available on our API! n2 is a very strong computer-use model for agentic and knowledge…
    @togethercomputeProud to continue to partner with @yutori_ai to power their Navigator product suite!

    13 Sources

    @theruiwangFrontier-level computer use: 65.2% on OSWorld 2.0 at $1.46 per task
    @varibandiNavigator N2 should be the default CUA model you try, but dont listen to me listen to the benchmarks! It also costs an order of magnitude less than Opus :) https://yutori.com/blog/introducing-n2
    @_krunalshah_We're launching Navigator n2, a computer-use model that connects with computers deeper than just graphical interface. We've put in a lot of ❤️ and 🤔 into making the model particularly shine on long-horizon computer-use tasks, go check it out at https://yutori.com/n2!
    @abhshkdzWhen we started Yutori, our research goal was to post-train the best browser-use models for powering web agentic products. We did that with Navigator n1 and n1.5; models that were pareto-dominant (at the time of release) in accuracy and cost for browser-based workflows. And browser-use models are basically computer-use models (screenshots in, actions out) optimized for driving web browsers (only), i.e. the data and action sets are constrained to browser control specifically. With today's release of Navigator n2, we've broadened our ambition and gone beyond the browser, towards a computer-use model that can operate full-desktop environments, interleaving graphical (including browsers!) and command-line interfaces, and even writing short snippets of code to complete tasks. And still pushing the accuracy-cost pareto, this time for computer use. General-purpose frontier models ofc have a place in the world, particularly for discovering use-cases that weren't possible say 6 months ago, but if you're looking to ship computer-use / knowledge work agents today at scale, n2 is a compelling option at a fraction of the cost. We spoke with @alexrkonrad to share the thinking behind n2.
    @skanda_vaidNavigator n2 is out! n2 is trained to use computers the "computer way" and is frontier level across all CUA benchmarks! A lot of cool learnings along the way that we're hoping to share soon!
    @svpinoThe issue I have with computer-use models today is that they spend too much time clicking through interfaces. You can die of boredom waiting for these tools to finish a simple task. n2 is a new 27B-parameter computer-use model that handles this differently. When solving a problem, the model chooses how to complete each step based on the fastest available option. The model works in 4 different ways: 1. Use a GUI to interact with an application 2. Write terminal commands to work with the operating system 3. Call an API to access a service directly 4. Write code to handle a task programmatically As n2 works through a task, it switches between these options, and the end result is a much faster solution! For example, to complete a specific workflow, the model may browse the web, open an application, edit one file, run a few commands, and call an API. This is key when you are building agents that work across different tools. You want these agents to be faster, less fragile, and be able to work for longer. You can test the API and interactive playground at http://yutori.com/n2. Thanks to the Yutori team for partnering with me on this post.
    @deviparikhRT @abhshkdz: Excited to release Navigator n2, now available on our API! n2 is a very strong computer-use model for agentic and knowledge…
    @togethercomputeProud to continue to partner with @yutori_ai to power their Navigator product suite!