NeoBERT is an open-source encoder with 250 million parameters, released by MILA’s chandar-lab under a permissive license on Hugging Face. This means there are no direct licensing fees associated with downloading or utilizing its weights for research or commercial purposes. In practical terms, the "cost" of using NeoBERT primarily revolves around infrastructure and inference expenses rather than paying for the model itself. The authors have intentionally designed it as an accessible, plug-and-play alternative to BERT/ModernBERT, which eliminates the need for extensive computational resources.

Due to its compact and optimized design (including FlashAttention, RMSNorm, and a 4,096-token context), NeoBERT can efficiently operate on a single modern GPU or even on powerful CPUs. This capability results in very low per-request costs, typically well under a fraction of a cent for every 1,000 tokens in self-hosted environments, depending on the hardware and usage. Managed service providers that offer NeoBERT via APIs generally price it similarly to other small to medium encoders, leading to API costs that are usually in the range of cents per million tokens. This makes NeoBERT one of the most economical choices for large-scale embedding, retrieval, and classification tasks.