The Canadian Unicorn Cohere recently unveiled “Command A”, the latest version of its flagship model. Specially designed, as its predecessors, to meet the needs of companies, this LLM of 111 billion parameters, which combines performance and energy efficiency, competes with leading models such as GPT-4O and Deepseek-V3.
One of the major assets of Command A for companies is its minimum hardware footprint. While most comparable models require up to 32 GPUs, Command A works effectively with only two A100 or H100 GPUs, which results in a significant reduction in costs and latency as well as a higher execution speed. In addition to a faster generation of the first token, it can thus generate up to 156 tokens/s, a flow rate 1.75 times greater than GPT-4O and 2.4 times higher than Deepseek-V3.
Command performance A
COHERE has evaluated the performance of Command A compared to those of GPT-4O and Deepseek-V3, on academic benchmarks: MMLU ((general knowledge) ,, Math, Ifeval (monitoring of instructions), intelligent agent tests (BFCL, Taubench) and coding benchmarks (MBPPPLUS, SQL, REPOQA).
His capacities in monitoring instructions, in coding, in particular in SQL, and on agent tasks surpass those of his competitors.
In human evaluation tests, Command A, which covers 23 main languages, has surpassed its competitors on several languages, especially in dialect Arabic, where it turned out to be more coherent and precise than GPT-4O and Deepseek-V3. This ability to adapt to local contexts represents a strategic asset for companies operating internationally.
Optimized capacities for companies
Unlike his predecessor, which supported a context length of 128,000 tokens, Command A has a context length of 256 tokens, which makes him able to analyze long business documents. It incorporates advanced features such as generation increased by recovery (RAG) with verifiable quotes and the use of secure agent tools.
It is particularly effective for:
-
Analysis and extraction of information from large financial reports;
-
HR policies management according to local specificities;
-
The verification and interpretation of complex legal regulations.
Thanks to a fluid integration with North, the COHERE IA agents platform, Command A allows companies to develop tailor -made AI solutions while maintaining a high level of safety and compliance.
Availability and pricing
Already available on the COHERE platform, with an upcoming support by the main cloud suppliers, Command A is offered at a cost of $ 2.50 per 1 million tokens at entry and $ 10.00 per 1 million tokens at output. It is also accessible for research on Hugging Face.