The Information Machine

🇨🇳 Another good model from China.

Rohan Paul Twitter · Rohan Paul (@rohanpaul_ai) · 2026-07-01

Chinese researchers released Agents-A1, a 35B parameter open-source agent model under Apache 2.0 that claims performance comparable to 1-trillion-parameter models by training on long action sequences averaging 45K tokens and using domain-specific teacher models for knowledge transfer.

Open original ↗

Appears in

Extraction

Topics: ai-agentschinese-aiopen-source-modelsefficient-aiknowledge-distillation

Claims

  • A 35B parameter agent model claims to match 1-trillion-parameter model performance by training on longer verified task trajectories rather than scaling up parameters.
  • Training data consists of long action records averaging 45K tokens, teaching the model complete multi-step work processes including search, tool use, error correction, and answer verification.
  • Specialist teacher models are trained separately for search, science, instruction following, and tool use, then their skills are transferred into a single student model.
  • Agents-A1 achieves strong results across long-task benchmarks covering search, science, coding, tool use, and instruction following.
  • The model weights are publicly available on Hugging Face under the Apache 2.0 license.

Key quotes

A 35B agent model claims 1T-model performance by thinking longer, not growing bigger.
The technique is proposing a cheaper way to make strong AI agents: teach them longer verified work habits, not just make them bigger.