silx-ai/QuasarV4-400B-1M

QuasarV4

QuasarV4 is the first-ever 400B non-transformer language model, built on top of a hybrid architecture combining Liquid Neural Networks and an advanced Memory Bank system.

QuasarV4 can handle context windows exceeding 1 million tokens (even more)

It is designed to overcome the limitations of transformer-based architectures by solving long-context inefficiencies and drastically reducing computational cost.

---

Model

• 400 Billion Parameters (MoE Architecture)
• 25 Billion Active Parameters
• Transformer-Free Design
• Powered by Liquid Neural Networks
• Integrated Long-Term Memory Bank
• Efficient Long-Context Understanding

---

Training Progress

• Current Progress: 1% trained on a target of 15 trillion tokens
• Architecture is fully operational and under active development
• Preliminary training shows promising results in context reasoning and memory recall

---

Why Quasar?

To create a scalable, efficient, and highly contextual language model architecture that eliminates the overhead of transformers while expanding the model’s ability to think, reason, and adapt in real time.

---

Architecture

[Input] → [Liquid Neural Core] → [Memory Bank Routing] → [Token Output]

• The Liquid Core adapts over time with dynamic temporal reasoning.
• The Memory Bank stores long-term facts, reducing the need for repeated context.
• Designed for extreme scalability and low compute per token.

---

Status

• QuasarV4 is currently under early-stage training
• Intended to run on limited compute while enabling frontier capabilities
• Results will be published progressively

---

📢 Contact

Built by SILX AI

For research collaboration or questions: [email protected]