What is Command A+
Command A+ is an open-source multimodal large model released by Cohere Labs. It adopts a mixture-of-experts architecture with a total parameter count of 218B and only 25B activated parameters, balancing strong performance with inference efficiency. The model supports a 128K ultra-long context and processing in 48 languages, natively integrating visual understanding, agent task execution, complex reasoning, and machine translation capabilities. It offers three quantization versions, and private deployment requires as little as 1 B200 or 2 H100 GPUs. The model can be freely used commercially and further developed, making it an ideal base model for enterprise-grade agent applications, retrieval-augmented generation, and multimodal AI applications.
Command A+'s Main Features
- Multimodal Agent: Natively supports tool calling and multi-step Agent tasks, with the strongest performance in tool use and autonomous execution.
- Visual Input: Supports joint image and text input, enabling image-text understanding and generation.
- Reasoning and Translation: Built-in reasoning chains and world-class translation capabilities, covering 48 languages including Chinese, English, and French.
- Long Context Processing: 128K input context and 64K output length, suitable for long-document RAG and complex workflows.
- Efficient Deployment: Private deployment requires only 1×B200 or 2×H100 GPUs, with latency reduced by 30% compared to Command A Reasoning.
Command A+'s Technical Principles
- MoE Architecture: Adopts a mixture-of-experts architecture, with a total parameter count of 218B and only 25B activated parameters, balancing performance and inference efficiency.
- Unified Weights: Integrates vision, reasoning, translation, and Agent capabilities into a single set of model weights, eliminating the need to stitch together multiple models.
- Native RAG Optimization: Forms an integrated retrieval-augmented generation stack with Cohere Embed v3 and Rerank 3.5, supports native citation tracing, and has a significantly lower hallucination rate than general-purpose models.
How to Use Command A+
-
Download Weights: Use the Hugging Face CLI tool to pull the BF16, FP8, or W4A4 quantized version of the model files locally.
-
Transformers Loading: Use the pipeline interface of the Hugging Face Transformers library to directly load the model for image-text inference.
-
vLLM Deployment: After installing vLLM, specify the model path to start the service, which automatically exposes an OpenAI-compatible API endpoint.
-
SGLang Deployment: After installing SGLang, start the inference service, which also provides an OpenAI-compatible API for external calls.
-
Docker Run: Execute a Docker command to launch the containerized service with one click, without manually configuring the dependency environment.
-
API Call: Send a standard Chat Completions request to the local service, and pass text and image URLs in the message body to trigger inference.
Command A+'s Core Advantages
-
Enterprise-Grade Efficiency: Throughput increased by 110% over the previous generation, latency reduced by 30%, designed specifically for large-scale production deployment.
-
Extremely Low Deployment Threshold: W4A4 quantization requires only 1 B200 or 2 H100 GPUs for private deployment.
-
Data Sovereignty: Supports local deployment and mainstream cloud platforms, meeting compliance requirements in finance, healthcare, and other sectors.
-
Open-Source Friendly: Apache 2.0 license, with weights, code, and fine-tuning tools fully open, and no restrictions for commercial use.
Command A+ project address
- Project official website：https://cohere.com/blog/command-a-plus
- HuggingFace model repository：https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16
Comparison of Command A+ with similar competitors
| Dimension | Command A+ | OpenAI GPT-4o |
|---|---|---|
| Developer | Cohere Labs | OpenAI |
| Open-source license | Apache 2.0(fully commercial) | Closed source / proprietary API |
| Model architecture | MoE（218B / 25B） | Dense (undisclosed) |
| Multimodal capabilities | ✅ Vision + text + tools | ✅ Vision + audio + text |
| Context window | 128K input / 64K output | 128K |
| Private deployment | ✅ 1×B200 or 2×H100 | ❌ API calls only |
| API price (per million tokens) | $2.50 / $10.00 | $2.50 / $10.00 |
| Native RAG ecosystem | ✅ (Embed + Rerank + generation) | ❌ Requires self-building |
| Native Agent tool calling | ✅ | ✅ |
| Enterprise-grade RAG stack | ✅ End-to-end | ❌ |
| Quantization support | BF16 / FP8 / W4A4 | Not applicable |
Command A+ application scenarios
-
Enterprise knowledge base Q&A: Based on a 128K long context and native RAG capabilities, it enables cross-document intelligent Q&A with precise citation tracing.
-
Multilingual translation and localization: Covers 48 languages, suitable for cross-border e-commerce, simultaneous interpretation at international conferences, and multilingual content production.
-
Visual Agent automation: Combines image understanding and tool calling to enable end-to-end automated workflows such as invoice recognition, receipt entry, and report parsing.
-
Private deployment for compliance industries: Institutions such as those in finance, healthcare, and government can deploy locally or in a private cloud to meet data sovereignty and regulatory compliance requirements.
-
High-throughput intelligent customer service: With the high throughput and low latency characteristics of the MoE architecture, it supports large-scale concurrent conversations and real-time retrieval-augmented responses.