Cohere AI Updates: May 21, 2026

1. Cohere releases Command A+, a 218B MoE under Apache 2.0 with native citations and 48-language tokenizer

Cohere. Cohere released Command A+, a 218 billion-parameter sparse mixture-of-experts model with 25 billion active parameters, a 128K input window, and 64K maximum generation length, distributed under an Apache 2.0 license on Hugging Face in 16-bit, 8-bit, and 4-bit quantizations and via Cohere’s Model Vault managed inference. The model targets sovereign critical infrastructure for enterprises and public-sector buyers, with native citation generation that emits explicit “grounding spans” linking factual claims to specific source documents or database rows, and a new tokenizer that natively supports 48 languages and cuts Arabic by 20%, Japanese by 18%, and Korean by 16% over Command A. Benchmark jumps versus Command A Reasoning include τ²-Bench Telecom rising from 37% to 85%, Terminal-Bench Hard agentic coding from 3% to 25%, agentic QA up 20 points, and spreadsheet analysis up 32 points, with MMMU Pro multimodal at 63%; Cohere claims the model runs on as little as two H100s or a single Blackwell GPU with imperceptible quality loss. Source