Coding agents have mostly lived on your laptop, bottlenecked by local compute and tied to your active terminal session. Today, they are moving to the cloud. With SAGE 2.4 Actus, agents can run on their own, spin up parallel tasks, and notify you only when the job is done.

You can start them directly from the RUNE CLI or right within the chat interface, offloading complex, multi-step coding tasks without ever breaking your flow.

Powering this paradigm shift is SAGE 2.4 Actus, the new default intelligence engine in RUNE. Built explicitly to run for long stretches on coding and productivity work, it forms the backbone of our new Work mode—a powerful agentic system designed for deep research, codebase analysis, and autonomous cross-tool actions.

Highlights.

1. Rebranded & Refined: Formerly SAGE-32B, the SAGE 2.4 Actus architecture sets a new standard for tool calling, error recovery, and step-by-step problem solving.
2. Strong real-world performance: Outperforms comparable models like Qwen2.5-32B and Llama-3-70B on rigorous agentic evaluations like AgentBench and IFEval.
3. Agentic Workflow Integration: Built specifically to run for long stretches on coding, analysis, and cross-tool actions, forming the default model for RUNE.

Model Capabilities.

SAGE 2.4 Actus leverages a highly refined 32-billion parameter foundation. Enriched through rigorous Iterative Distillation and inverse reasoning methodologies, it demonstrates unprecedented capability in complex step-by-step problem solving. This release represents our unwavering commitment to the Actus lineage—spearheading our research into truly autonomous, long-horizon planning models.

The model is a dense 32B architecture featuring a massive 256k token context window, natively capable of handling long-form instruction-following, heavy reasoning, and intricate repository-level coding all within a single set of weights. By optimizing the KV cache and implementing advanced RoPE scaling, SAGE 2.4 Actus maintains near-perfect retrieval accuracy even at the very limits of its context window, avoiding the "lost in the middle" phenomenon that plagues lesser models.

Furthermore, reasoning effort is now dynamically configurable per request. The very same model can rapidly answer a quick chat reply with extreme efficiency, or autonomously work through a complex agentic codebase refactor by enabling an extended inference loop. We have also trained the vision encoder entirely from scratch to handle variable image sizes and highly complex aspect ratios, allowing for native visual debugging and UI-to-code synthesis without requiring external OCR tools.

Use Cases.

• Automated Codebase Refactoring: With its 256k context window and long-horizon planning loop, Actus can digest thousands of lines of legacy code, design a modern architectural plan, and sequentially refactor an entire monolithic application into microservices overnight.
• Threat Intelligence Orchestration: Actus can autonomously monitor server logs, query internal security tools via APIs, identify potential breach vectors, and draft mitigation policies before an engineer ever steps in.
• End-to-End Testing Generation: Beyond just writing unit tests, Actus can spin up virtual browser environments, navigate UI flows, and construct comprehensive Cypress or Playwright test suites based purely on a figma design or a textual description.