Hugging Face 
GitHub Long-horizon, multi-tool execution
- Maintaining goals, states, and constraints across long execution chains.
- Coordinating retrieval, decision, execution, and verification tools.
- Adjusting its path when intermediate results deviate.
Occamy Team · 2026
Occamy
Evolving data for capable real-world agents.
Post-trained from Qwen3.6-35B-A3B for long-horizon decision-making, multi-tool collaboration, and complex business tasks. Through supervised fine-tuning and reinforcement learning, Occamy demonstrates execution stability and task-completion ability on claw-style evaluations that exceed what its parameter scale alone would suggest.
SFT + MIX-RL · Agentic Post-TrainingDATA RSI · Recursive Data Evolution
Evaluation notes
* indicates that the model was evaluated by our team.
Terminal-Bench 2.1: Harbor/Terminus-2 harness; 3h timeout, 32 CPU/48 GB RAM; temp=1.0, top_p=0.95, top_k=20, max_tokens=80K, 256K ctx; avg of 2 runs.
SWE-bench Pro: Internal agent scaffold (bash + str_replace_edit tools); temp=1.0, top_p=0.95, 200K ctx; unmodified public set; all models re-evaluated in-house under this identical setup, and no officially reported numbers are adopted in this table.
τ³-bench: We report Banking Knowledge results following the official example configuration, using AllTools retrieval, GPT‑5.2 (reasoning_effort=low) as the user simulator.
Expand complete benchmark table
01 · Overview
Built for long-horizon decisions and multi-tool collaboration
Occamy is a compact agent model post-trained from Qwen3.6-35B-A3B for long-horizon decisions, multi-tool collaboration, and complex business tasks. It coordinates tools, adapts to changing task states, follows constraints, handles branching workflows, and turns sequences of actions into verifiable business outcomes. Through supervised fine-tuning and reinforcement learning, it delivers stable execution and strong task-completion performance on claw-style evaluations—outperforming expectations for its size.
At the core of Occamy is Data RSI, a recursive self-evolution mechanism that connects data construction with model training. The Data Engine composes executable tasks from personas, tools, fixtures, skills, constraints, and graders. Generation and validation systems iteratively refine each task until it is solvable, executable, and reliably scored. After training, the model returns to these environments for rollouts, and its performance drives three data strategies: distilling transferable behavior patterns from mastered tasks, prioritizing tasks that are not yet consistently solved, and identifying capability gaps from persistent failures to guide new task generation.
Built on this loop, Occamy demonstrates a new path for scaling agent capabilities: evolving data and environments enable a compact model to handle complex real-world work.
02 · Capabilities
Core capabilities for real-world operational work
Complex business-task reasoning
- Reasoning over permissions, business rules, relationships, and state transitions.
- Handling cross-system prerequisites across operational workflows.
- Turning requests into complete, compliant, and verifiable outcomes.
Authentic real-world interaction
- Acting through real tools instead of isolated static prompts.
- Reading changing system states and observing each action’s consequences.
- Adapting to actual feedback, permissions, and constraints.
03 · Data RSI
Data RSI: A Self-Evolving Strategy for Agentic LLM Post-Training
As model capabilities improve, training data depreciates quickly: older tasks are soon mastered reliably, while problems truly beyond the capability frontier are often too complex to be used directly for model training.
The key data challenge is therefore not producing more samples, but continually answering one core question:
What should the model learn next to make the greatest progress?
Data RSI is our answer. The Data Engine constructs executable tasks from personas, tools, fixtures, skills, constraints, and graders, then iteratively validates and refines them. The Training Engine updates the model and evaluates it through environment rollouts; the resulting performance signals guide the next cycle of data generation.
Each model generation’s execution results help define the data for the generation that follows.
03.1
From a task to an executable world
An agent does not face an isolated question. It operates inside a digital environment containing entities, relationships, tools, and state transitions.
TASK = PERSONA × TOOL × FIXTURE × SKILL × CONSTRAINT × GRADER
For example, an East China customer-operations specialist asks the model to update a customer’s contact information and synchronize the associated order...
We decompose the task into six elements:
The customer operator’s identity, permissions, and information boundaries.
Tools for querying and updating customers and orders.
Two same-name customers, their initial states, and associated orders.
The Customer Profile Skill provides procedures for identity lookup and order updates.
Disambiguate by company, region, and order relationships without changing unrelated records.
Verify the target customer and order while confirming that no other object changed.
The model must not only select the correct customer, but also complete the associated update and check for side effects. Missing the order update or modifying the same-name customer does not count as task completion.
Changing the role, skill, permissions, initial state, or tool combination creates different execution paths for the same business goal. Data no longer comes from surface-level rewrites of old problems, but from continuously expanding combinations of environments.
03.2
Generate tasks—and verify them
A task is not finalized in a single generation step. Instead, the generation and validation systems refine it through repeated feedback cycles.
Returning to the same-name customer task, the generation system first constructs the instruction, environment, skills, tools, and grader. The validation system then executes the task against the reference solution and actively tests incorrect paths, such as selecting the wrong customer or omitting the associated-order update.
For tasks with complex or ambiguous business rules, human experts also participate in review, surfacing issues that automated validation may miss and providing targeted guidance for validation.
Validation may find that the available clues cannot uniquely identify the customer, that missing the associated-order update still passes, or that substituting the target customer is not detected by the grader. The generation system responds by adding clues, adjusting corresponding fixtures, and repairing the grader before returning the task for another validation pass.
The task must eventually satisfy three conditions:
Solvable
The correct path can be found using only visible information.
Executable
Correct operations produce the expected state changes.
Discriminative
Wrong targets, missing steps, and unintended modifications are rejected.
Only when the correct path passes reliably and incorrect paths cannot bypass validation does the task enter the training data pool for the next training cycle.
03.3
Keep data half a step ahead of the model
A task’s value depends on the current model’s capabilities. Tasks that are too easy provide little learning signal, while tasks that are too hard tend to produce ineffective exploration. The most valuable data usually lies where the model is close to learning, but has not yet achieved stable mastery.
Repeated rollouts over the same task family produce three data signals:
Mastered → Distill BehaviorsExtract behavior patterns from consistently successful trajectories and transfer them into new worlds and tool chains.
Learnable → PrioritizePrioritize tasks and variants with unstable performance to concentrate training near the current capability frontier.
Too Hard → Extract GapsLocate missing capabilities from persistent failures and use them to guide more targeted task generation.
In the same-name customer task, stable success can distill the pattern “retrieve candidates → resolve by relationships → update associated state → check side effects.” Occasionally missing the order update becomes a priority training signal. If disambiguation repeatedly fails, the system generates new tasks around multi-hop relational reasoning.
Together, these signals update the Task Policy and continuously locate the model’s capability frontier. As training progresses, Too Hard becomes Learnable, Learnable becomes Mastered, and mastered behaviors move into more complex worlds.
04 · Summary
A new path for agent scaling
Occamy explores an alternative to scaling parameter count alone. Starting from Qwen3.6-35B-A3B, SFT and mix-RL produce stronger execution stability and task completion across long-horizon decisions, multi-tool collaboration, and complex business operations.
Data RSI is the core of this process. Each model generation’s executions shape the next data cycle: mastered tasks distill transferable behaviors, learnable tasks are prioritized, and verified-solvable tasks that remain too hard expose the next capability gaps. Candidate tasks return to training only after real environment execution and grader-based validation. Data becomes a moving system rather than a static resource.
Occamy is still an early exploration. Broader business worlds, more complex tool relationships, stronger verification, and long-term stability in open environments remain open problems. We plan to release Occamy model weights and part of the Data RSI dataset so the community can reproduce, evaluate, and extend data-and-environment self-evolution for agents.
As models advance, data must evolve in step.Occamy is our first step in this direction.
05 · Citation
Cite Occamy
If the Occamy model or open dataset supports your work, please use the relevant citation below.
Model citation
@misc{occamy2026,
title = {Occamy: Evolving data for capable real-world agents.},
author = {{Occamy Team}},
year = {2026},
url = {https://huggingface.co/occamy-ai}
}Dataset citation
@misc{occamy_data2026,
title = {Occamy Data: Data RSI for Agent Training},
author = {{Occamy Team}},
year = {2026},
url = {https://huggingface.co/occamy-ai}
}