> For the complete documentation index, see [llms.txt](https://docs.gradient.network/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.gradient.network/the-open-intelligence-stack/echo.md).

# Echo

{% hint style="success" %}
Research paper: [published](https://arxiv.org/pdf/2508.05387)
{% endhint %}

{% hint style="success" %}
Models: [open-sourced](https://huggingface.co/GradientResearch/models)
{% endhint %}

As AI systems grow in scale and generality, RL has become a critical layer for alignment. However, most co-located RL frameworks face a deadlock: inference interrupts training, training delays inference.

Echo is Gradient’s distributed RL framework, running on everyday consumer devices.

## Dual Swarm Architecture

Echo breaks the cycle by decoupling them into two independent swarms:

* Inference on distributed devices (via [Parallax](/the-open-intelligence-stack/parallax.md))
* Training on high-performance clusters

<div align="left"><figure><img src="/files/5b9069V93VS25DC6ZEOb" alt="" width="563"><figcaption></figcaption></figure></div>

## Echo Modes

We have open-sourced several early explorations with Echo, including:&#x20;

* A 7B math model beating Qwen2.5–32B
* A 30B Sokoban model outperforming DeepSeek-R1 & GPT-OSS-120B
* A 32B LoRA logic model ≥99% accuracy&#x20;

See all models in our [Hugging Face Collection](https://huggingface.co/GradientResearch/models).

## Learn More

Read the [blog](https://gradient.network/blog/echo-distributed-reinforcement-learning).\
Read the full [research paper](https://arxiv.org/pdf/2508.05387).