[research] 35B agent matches 1T-parameter performance via horizon scaling

[github-actions[bot]]

🔬 The Finding

Shanghai AI Lab released Agents-A1, a 35B Mixture-of-Experts model that matches or beats trillion-parameter models (Kimi-K2.6, DeepSeek-V4-Pro) on long-horizon agent benchmarks. The key insight: instead of scaling parameters, they scaled agent horizons — training on 45K-token trajectories built from knowledge-action-observation chains, then distilling six specialized domain teachers into one deployable model.

⚙️ What It Means for Agentic Workflows

1. Smaller models + richer trajectories beat bigger models. If you're choosing a backbone for an automated workflow, a well-trained 35B model can outperform a 1T model on multi-step tasks — inference cost drops dramatically without sacrificing quality.

2. Trajectory quality is the new hyperparameter. Workflow designers should invest in building high-quality, long-horizon training trajectories (tool calls, observations, verifier feedback) rather than always reaching for a larger model. The data pipeline matters more than model size.

🔗 Source

Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent — June 29, 2026

Generated by Daily Agentic AI Research Digest · 39.1 AIC · ⌖ 12.1 AIC · ⊞ 24.2K · ◷

• expires on Jul 8, 2026, 10:46 AM UTC

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-07-08T10:46:19.652Z.

Closed by Workflow