Trillion-Parameter Performance From a 35B Agent? Agents-A1 Explained

What Scaling the Horizon is

Scaling the Horizon is a new way of thinking about how to raise AI performance. The mainstream view has been "make the model bigger to make it smarter," but this paper points to a different path. The philosophy of running long tasks on few parameters contrasts with the giant MoE in our guide to LongCat-2.0; reading them together makes the difference clear.

Scaling the horizon, not the parameters

The paper's claim is right there in the title. Rather than enlarging the model, it extends the agent's "horizon," so that a 35B MoE model reaches trillion-parameter-level performance.

"Horizon" here refers to how long a sequence the model can look ahead over as it acts. Instead of finishing in a single reply, it widens the span it can handle across a long process of researching, thinking, and taking the next step.

Training on trajectories averaging about 45K tokens

The concrete method for extending the horizon is how the training data is built. Agents-A1 is trained on long "knowledge-and-action trajectories" averaging about 45K tokens.

By learning entire long workflows rather than short one-off Q&A, the model learns to "act with the next steps in view." That's the aim: lifting practical performance without adding size.

What Agents-A1 is, and its release

Beyond the idea, the model itself is public. Here is who built it and how far you can use it.

Shanghai AI Lab's 35B MoE (arXiv 2606.30616, ~50 authors)

The creator is a leading research institution. The paper's author line reads "Agents-A1 Team Shanghai Artificial Intelligence Laboratory," identifying it as Shanghai AI Laboratory's work. The paper was posted as arXiv 2606.30616 on June 29, 2026, with about 50 authors, and drew attention on Hugging Face Daily Papers the next day.

Released under Apache 2.0, built on Qwen3.5

The terms are open too. The model weights are released under the Apache 2.0 license and can be used broadly, including commercially. It is built on Qwen3.5-35B-A3B, with context extended up to 262K tokens.

Because it's released under an open license, researchers and developers can readily verify it hands-on.

Performance and why extending the horizon matters

"35B rivaling a trillion" is a bold claim; here's what it means in practice.

Benchmarks on par with trillion-parameter models

The paper shows comparisons against models far larger in size. Agents-A1 is placed alongside trillion-parameter models Kimi-K2.6 and DeepSeek-V4, with reported results holding their own on long-horizon search (SEAL-0) and instruction following (IFBench). These figures come from the paper (arXiv 2606.30616) and make the case that size alone doesn't determine performance.

The point of running a small model for longer

What Agents-A1 shows is an option beyond "make it bigger": "train the ability to look further ahead." If a smaller model can reach practical performance by learning long workflows, it opens the door to useful agents that keep running costs down. The idea of using smaller models practically on your own hardware also connects to our practical local-LLM coding guide.

In short, Scaling the Horizon presents "a winning path apart from the parameter race." Because it's released openly, whether the idea holds up can be tested on real tasks.

When feeding long task material to an agent like this, converting it to Markdown first preserves structure and makes it easier to handle.

Free ToolURL to Markdown ConverterConvert any public web page URL to Markdown. Preserves headings, tables, lists, and links — perfect for LLM and RAG preprocessing, research notes, and archiving web articles.Try it now →