What changed

InternScience released the Apache-2.0 Agents-A1 35B-A3B checkpoint, selected evaluation code and a technical report for long-horizon search, engineering, scientific research, instruction following and tool use.

Why it matters

The release gives teams a downloadable specialist checkpoint and evaluation artifacts for testing whether a small-active-parameter MoE can support longer agent trajectories.

What to do

Pin the 35B-A3B revision, reproduce the published tasks, then test a private long-horizon set with strict token, tool and retry budgets. Record completion, intervention, cost and final-state correctness.

Caveat

The model's reported gains are author measurements, and selected evaluation code does not establish broad reliability. Long trajectories can hide compounding errors and require enough memory to load the full checkpoint.