Yutori's n2 runs your desktop on a budget

Most computer-use models treat desktop automation like an arm with a mouse attached: staring at pixels, hunting for buttons, and clicking one by one.

Yutori just launched Navigator n2, a 27B parameter model designed to interact with your computer the way a power user actually does.

Instead of restricting itself to human-oriented GUIs, n2 natively blends desktop clicks, terminal commands, tool calls, and custom code across Linux, macOS, and Windows. If a task requires renaming 100 images, it writes a quick shell loop rather than opening 100 context menus. When visual verification is actually needed, it seamlessly flips back to the GUI.

Smarter interaction over brute force

While Yutori's earlier n1 and n1.5 models focused strictly on browsers, n2 tackles the full operating system.

The benchmark numbers punch well above its 27B weight class. On OSWorld 2.0, which grades complex long-horizon workflows, n2 scored 65.2%, outpacing OpenAI's GPT-5.6 Sol at 62.6% and Anthropic's Claude Fable 5 at 66.1%, trailing only Claude Opus 5. On the verified OSWorld benchmark, it hit 85.3%, and scored 70.3% on WeaveBench, which specifically evaluates mixing GUI clicks with code execution.

Yutori attributes this performance to a self-improving training flywheel. Over the past two months, autonomous agents explored native apps to generate over 10,000 real-world tasks across productivity, software engineering, and scientific domains. Every task comes with programmatic or model-based verifiers that check the actual environment state rather than trusting the agent's summary.

To train long-horizon execution efficiently, Yutori combined supervised fine-tuning with reinforcement learning, using an on-policy self-distillation technique that reached peak performance four times faster than standard group rollouts.

Via API, n2 costs $0.50 per million input tokens, dropping to $0.05 when cached, with outputs priced at $4.00 per million.

Why it matters

We think Yutori is addressing the two biggest bottlenecks holding back desktop agents: inference cost and sheer execution inefficiency.

For developers trying to automate messy back-office operations, running frontier models on repetitive desktop screenshots has been financially impractical. At 27B parameters with cached inputs at five cents per million tokens, n2 finally brings multi-step desktop agency into an affordable tier. It is definitely worth testing for internal workflows that previously stalled on API bills.

Strategically, this model challenges the assumption that visual reasoning alone will win computer automation. By treating the terminal and code as first-class citizens alongside clicks, Yutori achieves far cleaner execution with fewer steps.

The clear caveat here is safety. Giving an autonomous 27B model broad shell execution rights across a local machine introduces substantial blast-radius risks. Teams deploying n2 will need tight sandboxing and strict permission boundaries before turning it loose on sensitive enterprise machines.