Command A+

Claims and evidence

• Released 20 May 2026 ● Primary source[3]page date [2]entry date; Availability [1]General: Release Date

• Status Available ● Primary source[4]Command table: command-a-plus-05-2026, Status Live [5]

• Successor of Command A EditorialEditorial link: Cohere calls Command A+ the last model in the Command A family and says it unifies the capabilities of the earlier Command models in one model. It is a new mixture-of-experts model, not a fine-tune of Command A. ● Primary source[3]opening paragraphs [2]opening paragraph

• Change · Architecture Sparse mixture-of-experts design with 218B total and 25B active parameters, where Command A Reasoning is a dense 111B model.Compared with Command A Reasoning ● Primary source[3] [1]Model Properties: Model Size
• Change · Modality Adds image input to a general Command A model; Command A accepts text only.Compared with Command A ● Primary source[4]Command table, Modality [1]Model Properties: Input modalities
• Change · Languages Language coverage grows from 23 to 48 languages, including all official EU languages.Compared with Command A ● Primary source[3] [2]Key Features; Technical Details
• Change · Licensing Weights are under Apache 2.0 and not gated, where Command A Reasoning's weights are a gated research release under the non-commercial CC-BY-NC licence.Compared with Command A Reasoning ● Primary source[5]Model Summary [6]Model Summary

• Input text, image ● Primary source[1]Model Properties: Input modalities [4]Command table, Modality [5]Model Details: Input
• Output text ● Primary source[1]Model Properties: Output modalities [5]Model Details: Output
• Feature Reasoning modeThinking can be enabled or disabled, with token budgets, per the docs. ● Primary source[1]What Can Command A+ Be Used For? [5]Usage [2]opening paragraph
• Feature Tool use ● Primary source[5]Tool Use Capabilities [3]Snapshot table
• Feature Multilingual48 languages, including all official EU languages. ● Primary source[2]Key Features; Technical Details [5]Model Details: Languages covered
• Open weights YesApache 2.0, not gated; published in BF16, FP8 and W4A4 quantizations. ● Primary source[5]Model Summary [1]Distribution and Technical Integration [3]opening paragraphs [7]Sovereign by Design
• Context window 128K tokensInput context; maximum output 64K tokens. ● Primary source[2]Technical Details [4]Command table, Context Length [5]Model Details: Context Length
• Parameters 218B total, 25B activeSparse mixture-of-experts with 128 experts, 8 active per token plus one shared expert (model card). The second launch post says 24 billion parameters are active per prompt. ● Primary source[1]Model Properties: Model Size [2]opening paragraph [5]Model Summary: Model Size
• Access API, Open-weights download, cloud partnerCloud partner: the models overview lists an Azure AI Foundry model id (coherelabs-command-a-plus-05-2026-w4a4) and N/A for Amazon Bedrock, SageMaker and Oracle. The launch post also offers deployment in Cohere Model Vault. ● Primary source[2]Availability [1]Distribution and Technical Integration [4]Using Command Models on Different Platforms: Azure AI Foundry

Lineage

Variants

No variants recorded in this record.

Related AI Radar coverage

• Announcing Cohere's Command A+Cohere · Changelog · 20 May 2026

All model releases from Cohere on AI Radar →