Octopus v4

Octopus v4 is Nexa AI's 3B-parameter master model for what the company calls a graph of language models, described in arXiv 2404.19296 (April 2024, era-level date used here). Instead of answering everything itself, Octopus v4 uses patented functional tokens to route each user query to the most appropriate specialist open-source model and to reformat the query for that model's strengths, activating roughly 10B parameters of specialist capacity in total.

Nexa's vendor-reported figures claim an MMLU score of 74.8, competitive among models at that scale. The model is open-sourced with GGUF quantized builds for on-device deployment on phones and laptops. One caution carried over from the line: Octopus v2 shipped under CC-BY-NC-4.0, so verify the v4 model card license before commercial use.