Apodex 1.1

Apodex 1.1 is a large-scale language model family developed by Apodex, designed to achieve what the creators term "working capability"—the ability to make sustained, verifiable progress toward complex, long-horizon objectives. Released in August 2026, the model is built around the thesis that agentic intelligence requires not just raw reasoning, but the ability to interact deeply with executable environments such as files, search engines, and code sandboxes.

The system's development is centered on two primary scaling dimensions: Environment Scaling, which expands the diversity and verifiability of tool-use environments, and Agentic Coordination Scaling, which trains the system to decompose tasks and manage asynchronous work. The architecture utilizes a Mixture-of-Experts (MoE) structure, with the flagship proprietary model containing approximately 397 billion parameters. A more compact version, Apodex 1.1 Mini, features 35B total parameters (with ~3B active per token) and is released with open weights under the Apache 2.0 license.

Performance evaluations place Apodex 1.1 in the leading performance band for professional work, finance, and scientific research. On specialized benchmarks such as FrontierFinance and APEX-Agents, the model demonstrates strong capabilities in following long agent trajectories and coordinating multi-agent workflows. The system is supported by AgentOS, a shared execution harness and operating layer that maintains task state, provenance, and failure recovery across different agents and tools.

Apodex 1.1 supports a native context window of 262,144 tokens, facilitating the processing of large multi-file codebases and extensive research documents in a single request. While the flagship remains proprietary and accessible via API, the open-weight Mini version—fine-tuned from the Qwen 3.5 architecture—provides a locally deployable alternative that retains significant agentic capabilities, particularly for coding and data analysis tasks.