Command A+ W4A4
Cohere's quantized Command A+ multimodal model.
Last verified:
What is Command A+ W4A4?
Command A+ W4A4 is Cohere Labs' 4-bit quantized (NVFP4) version of the Command A+ open source model, a 218B total / 25B active parameter MoE model for agentic, multilingual, reasoning-heavy enterprise tasks with vision input support and a 128K context window. It is the recommended quantization, offering superior speed and latency with a smaller hardware footprint (runs on a single B200 or 2x H100).
Command A+ W4A4 pricing
Pricing model: Free
Free open source weights under Apache 2.0 license
Command A+ W4A4 pros
- Runs on a single B200 GPU (or 2x H100) thanks to 4-bit weight/activation quantization applied to MoE experts
- Recommended quantization with superior speed, latency, and negligible benchmark quality differences vs BF16/FP8
- Apache 2.0 open source license with agentic, multilingual, reasoning, and vision capabilities and 128K context
- Quantization-Aware Distillation (QAD) preserves full-precision teacher quality
Command A+ W4A4 cons
- Requires vLLM >=0.25.0 plus Cohere's melody library (>=0.9.0) for serving with accurate response parsing
- Transformers requires installing from source repository with necessary changes for this model
- Quantization applies only to MoE experts; attention path stays full precision, limiting further compression
Frequently asked questions about Command A+ W4A4
What GPU hardware does the W4A4 quantization require?
Minimum 1 x B200 (Blackwell) or 2 x H100 (Hopper).
What is the model size and context length?
25B active parameters out of 218B total, with a 128K input context length.
Does the 4-bit quantization degrade quality?
Cohere states all three quantizations show negligible benchmark quality differences; QAD distillation closes the residual gap.