---
title: 'New Free Open-Source Chinese AI Long Cap 2.0 with 1 Million Token Memory'
source: 'https://youtube.com/watch?v=MzkDIcNb67k'
video_id: 'MzkDIcNb67k'
date: 2026-07-17
---

# New Free Open-Source Chinese AI Long Cap 2.0 with 1 Million Token Memory

> Source: [New Free Open-Source Chinese AI Long Cap 2.0 with 1 Million Token Memory](https://youtube.com/watch?v=MzkDIcNb67k)

## Summary

Long Cap 2.0 is a new free and open-source Chinese AI model developed by a food delivery company, not a tech giant. It features 1.6 trillion parameters with Mixture of Experts, a 1 million token context window, and outperforms GPT-5.5 on the SWE-bench pro benchmark.

### Key Points

- **Origin** [00:00] — Developed by a food delivery app, not a major tech company.
- **Anonymous Testing** [00:00] — Ran anonymously on open router for 2 months under the codename Alpha.
- **Model Architecture** [00:00] — 1.6 trillion parameters with Mixture of Experts, activating only needed parameters per task.
- **Context Window** [00:00] — 1 million tokens of context, allowing entire knowledge bases to be processed at once.
- **Performance** [00:00] — Scored 59.5 on SWE-bench pro, beating GPT-5.5 which scored 58.6.
- **License** [00:00] — Fully open-source under the MIT license, free for anyone to use and build upon.

### Conclusion

Long Cap 2.0 demonstrates that impressive AI advancements can come from unexpected sources, offering competitive performance and full open-source accessibility.

## Transcript

New free, plus open-source Chinese AI is insane. This new Chinese AI has 1 million tokens of memory for free. This new Chinese AI is called Long Cap 2.0, and it just came from a food delivery app, not a tech giant. It ran anonymously on open router for 2 months under the code name Alpha, climbing the charts with nobody knowing who built it. It has 1.6 trillion parameters, but only activates the ones it needs for each task, saving power and running faster. The real game-changer is 1 million tokens of context. Feed it an entire knowledge base, and it holds all of it in memory at once. On SWE bench pro, it scored 59.5, beating GPT 5.5 at 58.6, and it's fully open-source under the MIT license, free for anyone to build on.