Ternary weight quantization is a practical direction that deserves more attention from anyone building or deploying AI models. The core insight is simple: by restricting weights to just three values (+1, 0, -1), you can dramatically cut model size and inference cost while retaining more representational power than binary networks. That trade-off matters for real-world deployment, especially when you're trying to run models on edge devices or serve them at scale.
The research community has taken this seriously since at least 2016, when papers like Ternary Weight Networks established the theoretical foundation. Most work has focused on post-training quantization, train in full precision, then compress. That approach works, but it treats the training and inference stages as separate problems. The more interesting question is whether you can train natively in ternary from the start. That is where the project called Aigarth, developed by Qubic, makes a genuinely unusual claim: native ternary training using an evolutionary selection mechanism instead of gradient descent. The argument is that this produces models that represent uncertainty more naturally and remain adaptive after training, rather than freezing into fixed weights.
We are not in a position to rigorously evaluate that claim. But we can say that the combination of native ternary training and evolutionary optimization is not a mainstream research direction. A search of peer-reviewed literature does not turn up established papers on this specific combination. That does not mean it is invalid, but it does mean the burden of proof rests on the developers to demonstrate that their approach outperforms the well-studied post-training quantization methods. Until that evidence appears, the practical path for most teams remains the proven one: train in full precision, then quantize to ternary for inference.
What this means for you is straightforward. If you are building models that need to run efficiently on limited hardware, ternary quantization is worth exploring today. The tools and research are mature enough to apply. The evolutionary training approach is interesting as a research direction, but it is not yet a drop-in replacement for your current workflow. Keep an eye on it, but do not wait for it. The efficiency gains from ternary weights are real and available now.