Breaking

InternScience Releases 35B MoE Agent Model Agents-A1

June 30, 2026 at 14:45 EDT

Foundation Models
AI Agents
Open Source

Around June 29-30, 2026, InternScience released Agents-A1, a 35B-parameter MoE agent model designed for long-horizon tasks. It features a 256K context length with built-in agentic reasoning, and its weights are published under Apache-2.0 on Hugging Face and ModelScope.

June 30, 2026 · Shanghai AI Laboratory — InternScience

Agents-A1: Scaling the Horizon, Not the Parameters

An open-weight 35B-parameter MoE model — with just ~3B active at inference — claims to match or surpass trillion-parameter-class systems on long-horizon agentic tasks by lengthening the execution horizon instead of growing the model.

35B

total parameters (MoE)

~3B

active at inference

256K

token context length

Apache-2.0

open-weight license

The efficiency claim, drawn to scale

Active parameters at inference — a mid-size model competing with trillion-scale frontier systems.

~3B

Agents-A1
active params

~1,000B+

Trillion-class
frontier model

Takeaway: the horizon — not the parameter count — does the heavy lifting.

Reported benchmark scores

GAIA · Top

96.0

IFEval · SOTA

94.8

FrontierSci-Olympiad · SOTA

79.0

BrowseComp · 35B-best

75.5

SEAL-0 (search) · SOTA

56.4

HLE (with tools)

47.6

SciCode · 35B-best

44.3

FrontierSci-Research · SOTA

40.0

Early praise

Strong long-horizon scientific research
Robust multi-step reasoning
Stable tool calling across extended sessions

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.