endless-frontier/Fx-Work
Fx-Work is a 35.1 billion parameter multimodal language model developed by endless-frontier, built on Qwen3.6-35B-A3B. It is specifically post-trained with 20K tool-interaction trajectories to interpret complex work materials, reason over operational constraints, and utilize tools in realistic file-and-tool environments. The model excels at producing deliverables that meet explicit professional requirements, achieving strong benchmark results in occupational work tasks.
Loading preview...
Overview
Fx-Work is a 35.1 billion parameter multimodal language model from endless-frontier, designed for professional work within realistic file-and-tool environments. It is based on Qwen3.6-35B-A3B and has been extensively post-trained using 20,000 tool-interaction trajectories. These trajectories are derived from occupational work scenarios involving real-world artifacts, domain-specific knowledge, actionable requests, and detailed evaluation criteria. The training process incorporates execution-guided verification to ensure consistency and accuracy.
Key Capabilities
- Multimodal Interpretation: Processes and understands diverse work materials.
- Operational Reasoning: Reasons over constraints and requirements inherent in professional tasks.
- Tool Use: Integrates and utilizes tools effectively to accomplish tasks.
- Deliverable Production: Generates outputs that satisfy explicit professional standards.
- Extended Context: Supports a native context length of 262,144 tokens, with 128K tokens recommended for long-horizon professional work.
Performance
Fx-Work demonstrates strong performance, surpassing other models at or below the 35B scale across five reported metrics, including GDPval Rubric, GDPval Elo, APEX-Agents pass@1, and JobBench Main/Easy. It also outperforms DeepSeek-V4-Pro-Preview (1.6T) on four of these five metrics, highlighting its efficiency and capability in complex work-related benchmarks.
Good For
- Automating professional tasks requiring complex interpretation and tool use.
- Applications needing robust reasoning in file-and-tool environments.
- Generating high-quality, professionally compliant deliverables.