SupraLabs_Over the past few days, AxionLab and I have spent a lot of time discussing the future of our research, our development strategies, and how we want to approach the open-source AI space moving forward. As a result of these conversations, we have officially decided to leave the CompactAI organization and withdraw our models from their ecosystem.
This was not a sudden decision, but a necessary step. As we progressed, our core philosophies regarding development methods, project direction, and what true open-source work means simply moved too far apart. We realized that to build the kind of models we believe in, we needed a clean slate and full creative control over our pipeline.
Joining forces for what's next
Instead of splitting our focus, we are now channeling 100% of our energy, hardware, and time into SupraLabs. By bringing our separate efforts together into a single, cohesive team alongside Harley-ml, we are in a much stronger position to build something genuinely meaningful.
We are shifting away from short-term experiments and focusing heavily on rigorous, transparent architecture design. No shortcuts, no black boxes—just clean, well-documented, and verifiable open-source code.
Split Focus Different Objectives
100% Open Source Shared Hardware & Code
Current progress on the 50M Llama
To show you what we mean by moving forward, our new 50M parameter Llama model has already been training around the clock for the past few days on 20 Billion tokens of FineWeb-Edu. The pipeline is running flawlessly on our hardware, and the metrics are looking incredibly solid and stable:
{'loss': '3.081', 'grad_norm': '0.6257', 'learning_rate': '3.061e-05'}
Progress: 131,522 / 152,588 steps (~86% completed)
Unlike our previous sub-1M tests, a 50M footprint gives the network enough internal capacity to actually hold onto logic, maintain proper English grammar, and map out complex semantic spaces. Once the pre-training wraps up in a few hours, we will immediately transition into SFT (Supervised Fine-Tuning) and strict instruction alignment to optimize it for conversational tasks.
Final thought
Leaving an established server or group is always a tough call, but it is often the only way to ensure your work stays honest and aligned with your values. SupraLabs is now our definitive home. We are incredibly excited about the data, the code we are writing from scratch, and what this unified team can achieve.