Navigator n2: Frontier Computer Use at a Fraction of the Cost

ListenNavigator n2: A 27B Agent That Knows When to Skip the GUI11:55

Show notes

Yutori's Navigator n2 is a 27-billion-parameter computer-use model that drives full desktops across Linux, macOS, and Windows — clicking, running shell commands, calling tools, and writing code as needed. We dig into its two big bets: a computer-native interleaved action space and a recursive data pipeline where computer-use agents build the tasks, verifiers, and rollouts that train the next model. We also unpack the reported benchmarks, the aggressive pricing, and the claim that on-policy self-distillation hit peak performance four times faster than group-rollout RL. **Sources** - Navigator n2: Frontier Computer Use at a Fraction of the Cost (Yutori)

Subscribe to the channel: https://podcast.lumalabs.link/27a25f51ea81d41a8253f46c6133a35e/feed.xml

TL;DR

Yutori releases Navigator n2, a 27B general-purpose computer-use model that operates full desktop environments across Linux, macOS, and Windows — a generalization from Navigator n1/n1.5’s browser-only scope. n2 reports OSWorld 2.0 65.2 / OSWorld-Verified 85.3 / MyPCBench 82.6 / MacAgentBench 83.1 / WeaveBench 70.3, sitting at the current cost-accuracy Pareto frontier: it undercuts Claude Opus 5 (70.6 on OSWorld 2.0) by a few points but is priced at $0.50 /$ 4 per M input/output tokens ($0.05 cached). The load-bearing design bet is computer-native action space — the model chooses between GUI clicks, shell commands, tools, and code at each step rather than being locked to any one interface — and the load-bearing training bet is a recursive CUA-in-the-loop data pipeline where computer-use agents generate the tasks, verifiers, and rollouts that train the next generation of computer-use agents. The post also reports that on-policy self-distillation (OPSD) reached peak performance ~4× faster in wall-clock than group-rollout RL during their long-horizon training.

Key claims

Navigator n2 is a 27B model targeting general desktop computer use across Linux, macOS, and Windows — a scope shift from Navigator n1 and n1.5’s browser-only specialization [§Introduction].
OSWorld 2.0 65.2, OSWorld-Verified 85.3, MyPCBench 82.6, MacAgentBench 83.1, WeaveBench 70.3 — a five-benchmark suite spanning long-horizon workflows, open-ended desktop tasks, personalized computer use, macOS-specific tasks, and GUI+CLI+code interleaving [§n2 is frontier-level across computer-use benchmarks].
Against reported baselines: Claude Opus 5 leads OSWorld 2.0 (70.6) but n2 leads OSWorld-Verified (85.3 vs Claude Fable 5 85.0, Claude Sonnet 5 81.2, Kimi K3 84.8, GPT-5.6 Sol 83.0, Gemini 3.6 Flash 83.0) — n2 is the first entry in the wiki claiming a top or near-top OSWorld-Verified number at 27B open-scale [Detailed benchmark table].
API pricing: $0.50 / M input,$ 0.05 / M cached input, $4.00 / M output — positioned as substantially cheaper than the frontier closed models it benchmarks against [§Get started with n2, Table 1].
Computer-native action space: n2 is trained to interleave GUI clicks, CLI/shell commands, tool calls, and short code snippets at each step, rather than being constrained to any one interface [§Computer-native computer use].
Cited efficiency argument: renaming 100 files in the shell is one for loop versus 100 GUI rounds — an interface-mixing framing that motivates the multi-interface design against pure-GUI computer-use models [§Computer-native computer use].
Cited failure-mode framing (attributed to OSWorld 2.0 and WeaveBench analyses): today’s frontier models on GUI+programmatic tasks either persist on a suboptimal interface or bounce between interfaces without progress — n2 is trained to choose the right interface at the right time [§Computer-native computer use].
Recursive data-generation loop: computer-use agents (CUAs) participate throughout the training data loop — task creation (CUA rollouts explore what is feasible in each application), verifier authoring (programmatic checks + LLM-rubric checks over environment end-state), stress-testing (CUA rollouts surface false positives, false negatives, reward hacks, and edge cases), and post-training failure analysis (rollouts reveal capability gaps that feed the next task batch) [§Training n2].
Task scale: >10,000 tasks generated across hundreds of applications over ~2 months, spanning productivity, communication, business services, engineering, infrastructure, data analysis, creative production, and specialized scientific workflows [§Training n2].
Training regime: SFT for breadth + RL for depth, with continuous curriculum refresh — RL tasks that become reliably solvable are demoted to SFT while new failure modes are promoted to RL. Focus areas explicitly named: long-horizon planning and execution, dynamic interface selection, self-context compaction, “hundreds of steps” trajectories [§Training n2].
On-policy self-distillation (OPSD) reached peak performance 4× faster in wall-clock than group-rollout-based RL in Yutori’s experiments, and was “particularly efficient at learning from tasks with zero success under the current policy” — a load-bearing claim that OPSD is not just cheaper but expands the trainable frontier by working on failure-only tasks [§Training n2].
Yutori frames this as a recursive improvement flywheel: “better computer-use models help create and refine better training data, and better training data in turn produces better computer-use models” [§Training n2].

Method

Navigator n2 is a 27B multi-OS computer-use model post-trained from an unspecified base via a combination of SFT and RL. The distinguishing methodological piece is the data pipeline, not the model architecture.

The pipeline is a closed loop with three roles for CUAs. First, task creation: CUA rollouts explore application state spaces to determine what is actually feasible, grounding new task specs in observed application logic rather than in written documentation. Second, verifier co-generation: each task ships with a verifier that inspects the environment end-state (queries a database, diffs a file, checks a calendar invite) or, when the criterion is softer, an LLM-based rubric. Third, stress-testing: candidate task+verifier pairs are exercised via CUA rollouts to surface (a) false positives — wrong agent behavior scored as correct; (b) false negatives — correct behavior scored incorrect; (c) reward hacks — the outcome is achieved via unintended trajectories; (d) environment bugs — mocked applications with exploitable surfaces. Each surfaced failure becomes a regression test case, and the task, verifier, or environment is refined jointly.

Training combines SFT (breadth over applications, workflows, and interface types) with RL (depth on harder tasks). The RL set is continuously refreshed: reliably-solved tasks move down to SFT while newly surfaced failure modes move up to RL. Post-training rollouts are analyzed for capability gaps that feed the next batch of task generation — this is what Yutori calls the recursive improvement flywheel. On-policy self-distillation is used as an efficiency-and-frontier-expansion mechanism in the long-horizon setting.

Model output pricing structure (input $0.50/M, cached input$ 0.05/M, output $4/M) implies aggressive prompt-caching design in the serving stack, consistent with a computer-use agent trajectory pattern where the system/context prompt is stable across many turns.

Results

Headline numbers reported by Yutori in the detailed benchmark table (all Yutori-reported; not third-party verified):

OSWorld 2.0 (long-horizon real-world workflows): n2 65.2, vs Claude Opus 5 70.6, Claude Fable 5 66.1, Kimi K3 58.3, Claude Opus 4.8 55.7, GPT-5.6 Sol 62.6, GPT-5.6 Terra 50.2, GPT-5.6 Luna 45.6, GPT-5.5 47.5, Gemini 3.7 Flash 47.9, Gemini 3.6 Flash 33.8, Gemini 3.1 Pro 30.6, Muse Spark 1.1 47.3.
OSWorld-Verified (open-ended tasks across real desktop/web apps): n2 85.3, vs Claude Fable 5 85.0, Kimi K3 84.8, Claude Opus 4.8 83.4, GPT-5.6 Sol 83.0, Gemini 3.6 Flash 83.0, Claude Sonnet 5 81.2, Muse Spark 1.1 80.8, GPT-5.5 78.7, Gemini 3.5 Flash 78.4, Claude Opus 4.7 78.0, Gemini 3.1 Pro 76.2, GPT-5.4 75.0, Gemini 3.5 Flash-Lite 74.0, Claude Opus 4.6 72.7, Claude Sonnet 4.6 72.5, GPT-5.4 mini 72.1, Muse Glimmer-30B 65.9.
MyPCBench (personalized computer use across user’s apps/accounts/history): n2 82.6, vs Claude Opus 4.8 62.01, Claude Opus 4.6 58.2, GPT-5.6 Sol 55.4, GPT-5.6 Luna 55.4, Claude Sonnet 4.6 50.5, GPT-5.5 45.1, GPT-5.4 mini 23.9.
MacAgentBench (real macOS tasks across 25 apps, GUI+CLI): n2 83.1, vs Claude Opus 4.6 73.7, GPT-5.5 66.71, Gemini 3.1 Pro 63.3, GPT-5.4 60.7, Claude Opus 4.8 58.41.
WeaveBench (tasks requiring GUI+CLI+code interleaving): n2 70.3, vs Claude Opus 4.7 53.2, GPT-5.5 49.9, GPT-5.4 46.5, Gemini 3.1 Pro 22.3.

Method-side result: OPSD reached peak performance ~4× faster in wall-clock time than group-rollout-based RL in Yutori’s training experiments, and was “particularly efficient” on tasks with zero-success rollouts under the current policy [§Training n2, no numbers table].

No third-party benchmarking, no ablation table, and no released weights or checkpoints. All numbers are Yutori-reported through their API-only release.

Why it’s interesting

n2 is the first filed datapoint that puts a 27B open-scale-parameter model (though API-only, weights not released) at or near the top of the four-benchmark computer-use quadruple that the concept pages track — OSWorld 2.0 (65.2), OSWorld-Verified (85.3), MacAgentBench (83.1), WeaveBench (70.3) — beating filed frontier models like Kimi K3 (OSWorld-Verified 84.8, per Kimi K3 Technical Report — Open Frontier Intelligence) and reported GPT-5.6 Sol / Claude Opus 4.8 / Gemini 3.6 Flash numbers on the same suite. This complements OSGym: Scalable OS Infra for Computer Use Agents (OSWorld-89 Pass@1 34.5 with Qwen-2.5-VL 7B via OS-replica RL) at the higher-capability end — same “post-training on full-desktop OS environments” thesis, an order of magnitude bigger model, an order of magnitude better OSWorld number. It also sharpens Kimi K2.5: Visual Agentic Intelligence‘s OSWorld-Verified 63.3 by re-anchoring the frontier upward.

Three connections matter for our own work. First, the computer-native interleaved action space is a stronger claim than the “GUI vs API tools” split that Tool-Use Agents currently frames — n2’s design bet is that GUI, CLI, tools, and code are one action space, and the model’s job is dynamic interface selection, not primarily execution. This complements Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World-Centric Foundation GUI Agents (Qwen-UI-Agent’s unification of mobile/desktop/web/deep-research) at 27B with a shared design principle. Second, the recursive data-generation loop — CUA-driven task creation, verifier co-generation, stress-testing, and post-training gap analysis — is a concrete instance of the “environment-first” pattern also seen in AgentOmnia: Scaling Agentic Models for Full-Scenario Applications (AgentOmnia’s 5,018 environments with programmatic verifiers) and PostTrainBench: Can LLM Agents Automate LLM Post-Training? (PostTrainBench’s LLM-agent-automates-training story). Third, the OPSD 4× wall-clock claim is a new datapoint for the OPSD-vs-GRPO comparison that AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning, Self-Distilled Agentic Reinforcement Learning, and Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe have been circling — the specific empirical claim (OPSD is “particularly efficient at learning from tasks with zero success under the current policy”) aligns with AgentOPSD’s Bayesian turn-level credit story on why OPSD-like teacher-student gaps can supervise trajectories that GRPO’s outcome reward cannot.