AI BriefingAgents-A1 Released, Achieving 1T-Class Performance with a 35B Model
Key point
InternScience released Agents-A1, a 35B MoE agent model that rivals 1T-class models by scaling agent horizon instead of parameters.
Details
Agents-A1 is a 35B-scale Mixture-of-Experts agent model released by InternScience, built on Qwen3.5-35B-A3B and activating only about 3B parameters per token.
The core idea is horizon scaling instead of parameter scaling. The training data consists of long trajectories averaging 45K tokens, using the entire process from knowledge acquisition → tool calling → observation → verification as a supervisory signal.
Key results:
- SEAL-0 (retrieval): 56.4
- IFBench (instruction following): 80.6
- HiPhO (physics olympiad): 46.4
- FrontierScience-Olympiad: 79.0
- BrowseComp: 75.5
- HLE (tool use): 47.6
On multiple benchmarks, it matched or outperformed 1T-class models such as Kimi-K2.6, DeepSeek-V4-pro, and GPT-5.5. The core message is that it achieved this performance with a model roughly 30x smaller in total parameters.
The methodology rests on two pillars: (1) building a Knowledge-Action Graph (KAG) that connects external knowledge, actions, observations, and verification signals, and (2) a 3-stage training recipe that unifies 6 heterogeneous domains into a single model.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Jevper Released: A Jev Classification Wrapper for OpenAI-Compatible Models with Automatic Logprobs Detection