Cohere, the Canadian AI lab co-founded by former Google researcher Aidan Gomez, has unveiled Command A+, a groundbreaking 218-billion-parameter language model that redefines efficiency and accessibility in enterprise AI. Unlike proprietary alternatives, this model is now available under the permissive Apache 2.0 license on Hugging Face, marking Cohere’s first open-source release. The move underscores the company’s commitment to "sovereign AI"—empowering organizations to run advanced AI systems within their own secure infrastructures without sacrificing performance.
A leap in architectural efficiency
Command A+ stands apart with its Sparse Mixture-of-Experts (MoE) Transformer design. While the model contains 218 billion parameters, only 25 billion are active during any single inference step. This sparse activation allows the model to deliver the reasoning depth of a much larger system while maintaining the speed and resource efficiency of a smaller one.
The model’s breakthrough, however, lies in its quantization strategy. Cohere has optimized Command A+ for three precision formats: 16-bit (BF16), 8-bit (FP8), and a highly compressed 4-bit (W4A4) variant. The 4-bit quantization is particularly notable because it achieves near-lossless compression by targeting only the MoE experts while preserving full precision in critical attention pathways. This technique, combined with Quantization-Aware Distillation, minimizes performance degradation typically associated with aggressive model compression.
The efficiency gains are substantial. In benchmarks, the W4A4 quantized version delivers 375 tokens per second (TOPS) with a Time-to-First-Token (TTFT) latency of just 113 milliseconds. Compared to the previous Command A Reasoning model, this represents a 63% increase in output speed and a 17% reduction in latency. For enterprises, these metrics translate to faster response times and lower operational costs, especially in high-concurrency production environments.
Multilingual optimization and global readiness
Command A+ introduces a tokenizer refined for enterprise-scale multilingual applications. Supporting 48 languages, it significantly improves tokenization efficiency for non-European scripts. For instance, Arabic text requires 20% fewer tokens, Japanese 18%, and Korean 16% compared to previous models. Since inference costs scale with token usage, these improvements directly reduce expenses for global deployments, particularly in regions where language diversity is high.
The tokenizer’s enhancements also extend to specialized domains. By reducing fragmentation in input processing, the model maintains coherence in technical, legal, and financial documents, making it ideal for industries requiring precise, context-aware responses.
Native citations and agentic workflows
Beyond raw performance, Command A+ is engineered for autonomous and semi-autonomous workflows. It supports native citations, enabling the model to ground its responses in verifiable sources—a critical feature for industries like healthcare, legal services, and enterprise decision-making where transparency is essential.
Benchmarks highlight its capabilities in complex reasoning and specialized tasks. On 𝜏²-Bench Telecom, which evaluates intricate problem-solving, Command A+ achieved an 85% score, a dramatic leap from the previous generation’s 37%. In agentic coding, measured by Terminal-Bench Hard, the model improved from 3% to 25%. For advanced mathematics, it scored 90% on AIME 25, outperforming many larger proprietary models.
However, the model still trails leaders like DeepSeek V4 Pro in deep agentic coding and general intelligence tasks. Cohere acknowledges these gaps and positions Command A+ as a specialized tool optimized for enterprises prioritizing control, efficiency, and multilingual support over sheer scale.
The future of open, sovereign AI
Command A+ represents a strategic shift in the AI landscape. By releasing a frontier-grade model under an open license, Cohere is challenging the dominance of closed, proprietary systems. The model’s efficiency, combined with its native citations and multilingual support, makes it a compelling choice for enterprises seeking to deploy AI without vendor lock-in or performance compromises.
As the demand for sovereign AI grows, Command A+ could set a new standard for enterprise-grade language models. Whether it will disrupt the dominance of U.S. and Chinese giants remains to be seen, but its arrival signals a maturing open-source ecosystem where performance, accessibility, and control converge.
AI summary
Cohere’in yeni Command A+ modeli, Apache 2.0 lisansıyla sunulan ilk açık kaynaklı dev yapay zeka aracı. 218 milyar parametreye rağmen 25 milyar aktif parametreyle çalışan model, kayıpsız 4-bit sıkıştırma ve çok dilli destek sunuyor.
