Shanghai AI Lab Releases Sage-A1.5: A 35B MoE Agentic Model Advancing Long-Horizon Software and Scientific Discovery

Background and the Problem

As autonomous coding and scientific discovery agents tackle real-world complexity, the industry has often defaulted to brute-force parameter scaling, deploying 100B+ to 400B+ dense models. However, empirical studies reveal that agentic failure modes stem not from deficient linguistic reasoning, but from goal drift and cognitive fatigue over extended execution horizons. After dozens of tool invocations and terminal iterations, models lose track of initial objectives. Establishing effective horizon scaling—maximizing execution endurance within an economical parameter envelope—is a vital frontier. In arXiv:2610.28578, the Shanghai Artificial Intelligence Laboratory unveiled Sage-A1.5.

Architecture and How It Works

1. 35B Mixture-of-Experts Horizon Topology:

• Spans 35 billion total parameters with sparse activation, enabling cost-effective deployment on single-node workstation setups;
• Advances the 'Horizon Scaling' hypothesis, prioritizing long-trajectory state consistency over sheer parameter bloat;

1. Adaptive Capability Composition:

• Trains the agent not merely to trigger discrete APIs, but to dynamically compose foundational skills—code dependency parsing, architectural abstraction, and hypothesis verification—into fluid multi-turn workflows;

1. Repository-Level Alignment Corpus:

• Engineered on authentic repository snapshots with complex multi-file Git diffs, environmental dependencies, and noisy compilation logs.

Benchmarks and Measured Results

Evaluated across rigorous software engineering and scientific inquiry benchmarks (Figure 1, Section 4):

• Repository-Level Issue Resolution:
• Matches or exceeds dense 70B+ open models on multi-file issue resolution, maintaining strict compliance with project styling and test assertions;
• Long-Horizon Trajectory Stability:
• On tasks requiring over 50 continuous action steps, Sage-A1.5 reduces goal drift by over 45% relative to the original A1 model;
• Scientific Exploration Validation:
• Seamlessly conducts end-to-end scientific cycles—from literature cross-referencing to executable empirical Python simulations—with rigorous intermediate verification.

Getting Started for Developers

• Technical Report: Published on arXiv:2610.28578;
• Model Card & Hugging Face Checkpoints: Accessible via the Sage-A1.5 Hugging Face repository;
• Deployment Strategy: Research labs and enterprises operating under resource constraints (e.g., workstations with 2 to 4 GPUs) can deploy Sage-A1.5 as an air-gapped coding and scientific research copilot, eliminating recurring API token overhead.