01What is Agent A1 and How It Works

Agent A1 is a new free Chinese AI model that you can run locally on your own computer. It is a 35 billion parameter mixture of experts model, which means it only activates 3 billion parameters at a time to work efficiently. This model came from InternScience, a lab in Shanghai, and it is designed for agentic tasks like coding, tool use, and long-horizon research.

In this episode, Julian Goldie tested Agent A1 inside his agent factory, plugged into free cloud code with a local model. He showed how it can create landing pages, mini apps, and even a 3D solar system visualization. The model can run offline, so you can use it on a plane without internet.

It also has a 256K context window, which is enough for most tasks.

02Main Features of Agent A1

Agent A1 has several key features that make it useful for local AI work. Here are the main ones from the episode:

  • Free and local: You can run it fully free on your own machine, with no cloud costs.
  • Mixture of experts: It has 35 billion parameters but activates only 3 billion at a time, saving resources.
  • Agentic design: It is optimized for long-horizon tasks, tool calling, instruction following, and scientific research.
  • Fast speed: It ran at 95 tokens per second on a Mac Studio M4, faster than Jemma 4 Coda.
  • Long context: It supports a 256K context window, enough for most tasks.
  • Voice control: You can control it with your voice in the agent factory.
  • Benchmark leader: It ranks number two on the local leaderboard, beating GMF4.

These features make it a strong choice for local coding and research tasks.

03How to Set Up Agent A1 Locally

Setting up Agent A1 locally is straightforward. Julian showed a few commands and steps to get it running. Here is how you do it:

  1. Install Ollama on your computer if you have not already.
  2. Run Agent A1 model using the simple command provided in the episode (the exact command is not spelled out, but you can use standard Ollama pull commands).
  3. If you want to use it with Hermes, open your terminal and switch the Hermes model to Olama.
  4. Make sure Agent A1 is set up in your Ollama list.
  5. Create a new profile in Hermes and select the Hermes agent.
  6. Test it by sending a message to confirm it replies.

You can also plug it into your agent factory or other local coding tools. The setup is easy and quick, as Julian noted that applying the model is fast.

04How to Use Agent A1 for Coding and Creation

Using Agent A1 is simple. In the episode, Julian walked through several examples. He used the agent factory to give tasks like "build a keyboard." The model then built a keyboard with features like volume control and an oscillator switch.

He also showed other creations: a landing page, a neon snake game, and a 3D solar system visualization. These are basic mini apps, not frontier-level software like Fable 5, but they are impressive for a local model. You can control the model with your voice using the agent factory, and everything you build gets saved to your workspace.

For coding, you can use it directly in your terminal. It works well as a coding agent when plugged into tools like Hermes. It is also designed for long-horizon tasks, so you can give it complex projects that require multiple steps.

05Agent A1 Performance on Benchmarks and Tests

Julian tested Agent A1 against other local models and shared the results. On the local leaderboard, Agent A1 ranks number two. It is not as good as Quable, but it outperformed GMF4 in his tests.

  • Speed: Agent A1 ran at 95 tokens per second on a Mac Studio M4, faster than Quable and Jemma 4 Coda, which were much slower.
  • Score: On the local leaderboard, Agent A1 scored 4.8 out of 10, while Quable scored 7.14. Jemma 4 scored lower than Agent A1.
  • Benchmarks: On HLE, it outperforms Quen 3.6 and Step 3.5 Flash. Kimi, DeepSeq, and Chachipiti are not far off. It is particularly strong on science benchmarks.
  • Claims: The project page claims state-of-the-art on SEAL, zero long-horizon search, IFBench instruction following, and BrowseComp.

You can check GoldieBench.com to see more details and builds.

06Tips, Limits, and Mistakes to Avoid with Agent A1

Agent A1 has clear strengths and weaknesses. Julian pointed out that it is not good for visual tasks like UI design. It is more for research and agentic work.

Here are some tips and warnings:

  • Use it for science and search: It is fine-tuned for search tools and science, so use it for research tasks, not for pretty visuals.
  • Keep expectations realistic: It is not a frontier model, so do not expect it to compete with models like Fable 5 for high-end tasks.
  • Speed varies: Its reply speed depends on your setup and what is running in the background. It may not be super fast if your computer is busy.
  • Check your local setup: Make sure you have enough memory and processing power to run it efficiently.
  • Avoid heavy visual work: If you need a model for UI design, look elsewhere, like Quable, which is stronger but slower.

Following these tips will help you get the best out of Agent A1.

07Who Should Use Agent A1 and Why It Matters

Agent A1 is for anyone who wants a free, private, offline AI model for coding and research. It is especially useful for developers who work on local agentic loops, tool calling, and long-horizon tasks. Julian highlighted that you can run it on a plane with no internet, which is perfect for travelers or privacy-conscious users.

It matters because it makes advanced AI accessible without cost or cloud dependence. You can build mini apps, landing pages, and small games entirely locally. This is a big deal for hobbyists and professionals who want to experiment or prototype without spending money.

While it is not the best for visuals, its speed and agentic capabilities make it a strong tool for the right tasks.