Cohere Cracks Lossless Quantization and Native Citations with First Full Apache 2.0 Licensed Open Model Command A+

Why It Matters

By combining frontier‑grade reasoning with true open‑source licensing, Cohere gives companies the ability to run high‑performance AI securely in‑house, cutting costs and eliminating vendor lock‑in.

Key Takeaways

• Command A+ 218B model, only 25B active per token
• 4‑bit quantization runs on single Blackwell B200 GPU
• Apache 2.0 license enables unrestricted commercial use
• Tokenizer cuts Arabic tokens 20%, Japanese 18%, Korean 16%
• Native citations provide traceable outputs for regulated industries

Pulse Analysis

Cohere’s Command A+ represents a technical leap in large‑language‑model design. By adopting a sparse Mixture‑of‑Experts architecture and quantizing only the expert layers to 4‑bit while preserving full‑precision attention pathways, the model achieves near‑lossless compression. This enables inference on a single Blackwell B200 or two H100 GPUs, delivering up to 375 tokens per second and a 113 ms time‑to‑first‑token, dramatically lowering compute costs compared with trillion‑parameter rivals.

Equally transformative is the decision to release Command A+ under an Apache 2.0 license. Unlike Cohere’s earlier CC‑BY‑NC models, the new license removes commercial restrictions, allowing any organization to download, modify, fine‑tune, and deploy the weights without royalty fees. This open‑source stance fuels the sovereign‑AI movement, where enterprises can keep sensitive data on‑premise, avoid API pricing volatility, and retain full control over model updates and security patches.

For businesses, the model’s built‑in multilingual tokenizer and native citation capability address two persistent pain points: cost‑effective global deployment and regulatory compliance. Token reductions of up to 20% for non‑European languages translate directly into lower inference spend, while automatic grounding of factual claims mitigates hallucination risks in finance, healthcare, and legal workflows. As a result, Command A+ is poised to accelerate adoption of autonomous, tool‑augmented agents across sectors that demand both high performance and strict data governance.