Cohere launches Command A+: efficient MoE model for small teams

In a Reddit post, Cohere co-founder Nick Frosst unveiled Command A+, the company's first Mixture-of-Experts (MoE) model. This release follows months of behind-the-scenes work hinted at by co-founder Aidan Gomez. Command A+ prioritizes efficiency and accessibility over raw benchmarks. Thanks to innovative quantization techniques, the model runs smoothly on just one or two GPUs, making it feasible for smaller teams and individual developers to deploy powerful AI agents. The model is open-sourced under Apache 2.0, granting near-unrestricted access for commercial and research use. Frosst emphasized that while there is room for improvement in top-line performance, Command A+ is already one of the fastest and most responsive models in its category. Cohere remains enterprise-first, but actively values community feedback, which influenced this release after earlier critiques. The model files and details are available on Cohere's blog.

Command A+ marks a strategic shift for Cohere, blending enterprise-grade reliability with open-source flexibility. By opting for MoE architecture, the model activates only relevant parameters per task, boosting speed and reducing computational cost. Combined with aggressive quantization, it achieves high efficiency without sacrificing practical capability. This makes it ideal for building agentic systems—AI that can reason and take actions—particularly for teams with limited hardware. Cohere's commitment to Apache 2.0 licensing underscores a bet on community innovation: developers can fine-tune, integrate, and distribute without restrictions. The move also positions Cohere against larger rivals like Meta's Llama and Mistral, who have similarly embraced open weights. Early reactions from the community are positive, with users praising the model's responsiveness and low resource footprint. As Frosst noted, the feedback loop from open-source users directly influences Cohere's product roadmap, ensuring that Command A+ evolves with real-world needs.