Developer
Cohere

Command A+

Cohere's first Mixture-of-Experts model, released May 2026. 218B total / 25B active parameters, vision input, reasoning, and agentic tool use under Apache 2.0.

Released
May 20, 2026
Type
Mixture-of-Experts Language Model
Context window
128K tokens
Knowledge cutoff
April 1, 2025
License
Apache 2.0
🌐 Official WebsiteπŸ“š DocumentationπŸ’° Pricing & API
On this page

Overview

Command A+ is Cohere's enterprise flagship, released on May 20, 2026. It is the company's first Mixture-of-Experts model β€” 218 billion total parameters with 25 billion active per token β€” and the first Command model to fold vision input, reasoning, agentic tool use, and translation into a single set of weights.

Cohere positions it as "the last model in the Command A family, while being Cohere's first Mixture of Experts model." The design goal is deployability: Command A+ fits on 1Γ— NVIDIA B200 or 2Γ— H100s at W4A4 quantization, which is what makes a 218B-parameter model practical for on-premises and VPC deployments.

It is released under Apache 2.0, with weights published by Cohere Labs on Hugging Face.

A note on naming: Cohere has never released a model called "Command 2.0," and has never used an N.0 numbering scheme for the Command family. The real lineage is Command β†’ Command R β†’ Command R+ β†’ Command A β†’ Command A+.

Capabilities

  • Sparse MoE efficiency: 218B total parameters, 25B active. Cohere reports up to 63% higher output tokens per second and up to 17% lower time to first token versus Command A Reasoning at the same quantization and concurrency.
  • Native vision input: Accepts text and images; outputs text, reasoning, and tool calls.
  • Configurable reasoning: Thinking can be enabled or disabled per request, with a token budget to cap reasoning spend before the final response.
  • Agentic tool use: Tuned for multi-step tool orchestration; served through Cohere's Chat endpoint.
  • 48 languages: Coverage expanded from 23 to 48 languages, including all official EU languages, with a more efficient tokenizer for Arabic, Japanese, and Korean.
  • Translation: Cohere reports gains in machine translation and multilingual reasoning, and names WMT24++ as the public benchmark it evaluated with xCOMET-XL. It publishes no numeric translation score.

Technical Specifications

  • API model ID: command-a-plus-05-2026
  • Architecture: Sparse Mixture-of-Experts β€” Cohere's first
  • Parameters: 218B total / 25B active per token
  • Context window: 128K tokens
  • Max output: 64K tokens
  • Input modalities: Text, image
  • Output modalities: Text (Cohere's model card lists text only; reasoning traces and tool calls are surfaced as structured fields on the Chat response)
  • Knowledge cutoff: April 1, 2025
  • Languages: 48
  • License: Apache 2.0 (open weights)
  • Deployment footprint: 1Γ— B200 @ W4A4, or 2Γ— H100 @ W4A4
  • Endpoint: Chat

Weights

Published by Cohere Labs on Hugging Face. The headline checkpoint is CohereLabs/command-a-plus-05-2026-w4a4; BF16 and FP8 variants are also available.

Relationship to Command A

command-a-03-2025 remains live alongside Command A+. The two are not strictly ordered: Command A+ is faster, stronger on reasoning and agentic work, and multimodal, but it has a smaller context window β€” 128K versus Command A's 256K. Command A, in turn, caps generation at 8K tokens where Command A+ allows 64K. Pick by whether your workload is bounded by input length or by reasoning depth.

Use Cases

  • Enterprise agents: Multi-step workflows across internal tools, where tool-call reliability matters more than raw context length.
  • Private and on-premises deployment: Apache 2.0 weights on a single B200 make air-gapped and VPC deployments tractable.
  • Retrieval-augmented generation: Grounded question answering over enterprise document sets, a long-standing focus of the Command line.
  • Multilingual operations: Support, translation, and document processing across 48 languages, including full EU-language coverage for European deployments.
  • Document and chart understanding: Vision input for scanned documents, screenshots, and figures.
  • Agentic coding: Terminal-driven software tasks, where Cohere reports the largest generational jump.

Performance / Benchmarks

All figures below are reported by Cohere in the Command A+ announcement. Note that Cohere uses two different baselines depending on the benchmark β€” agentic results are measured against Command A Reasoning, multimodal results against Command A Vision. Conflating them overstates the multimodal gain.

Agentic (baseline: Command A Reasoning)

BenchmarkCommand A+Command A Reasoning
τ²-Bench Telecom85%37%
Terminal-Bench Hard25%3%

Multimodal (baseline: Command A Vision)

BenchmarkCommand A+Command A Vision
MMMU75.1%65.3%
MathVista80.6%73.5%
CharXiv Reasoning52.7%46.9%
MMMU Pro63%β€”

Cohere also reports a score of 37 on the Artificial Analysis Intelligence Index (a third-party aggregate), stating that this outperforms other leading open models. That figure is a launch-day number: Artificial Analysis's own article of May 20, 2026 says Command A+ "achieves 37 on the Artificial Analysis Intelligence Index" and names no index version. Artificial Analysis's live model page now reads "Command A+ scores 23 on the Artificial Analysis Intelligence Index," under Intelligence Index v4.1 β€” dated "June 2026β€”current" in Artificial Analysis's version history, and therefore published after Cohere's May 20 launch. It aggregates nine evaluations β€” GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. The two numbers come from different index versions and are not comparable, so 37 should not be quoted as a current score.

On efficiency, at matched quantization and concurrency Cohere measures up to 63% higher output tokens per second and up to 17% lower time to first token relative to Command A Reasoning.

For translation, Cohere names WMT24++ as its public benchmark, evaluated with xCOMET-XL. It reports no numeric score, so none is given here.

Limitations

  • Smaller context than Command A: 128K versus 256K tokens. Workloads that depend on very long single-prompt inputs may still be better served by command-a-03-2025.
  • No published token pricing: Cohere lists no per-token rate for Command A+. It is free until rate limits, after which capacity is bought as Model Vault instances rather than per token β€” which makes cost forecasting a capacity-planning exercise, not a token-count one.
  • Rate limits are the real ceiling: Because the free tier is bounded by rate limits rather than by spend, throughput planning matters more here than with a metered API.
  • Hardware floor: Even at W4A4, a B200 or a pair of H100s is the entry point for self-hosting. This is not a laptop model.
  • Knowledge cutoff of April 1, 2025: Over a year stale as of mid-2026; pair with retrieval or web search for current information.
  • Final model in its family: Cohere describes A+ as the last Command A model, so future improvements will presumably arrive under a new family name.

Pricing & Access

Cohere has not published per-token pricing for Command A+, and the model is not metered that way. Its documentation states that "for both trial keys and production keys, Command A+ is free until rate limits are reached." Past those limits, production runs on Model Vault, Cohere's dedicated managed deployment, whose hourly and monthly instance rates appear on the public pricing page β€” that page's per-token table covers only legacy models (Command, Command R, Command R+, Aya Expanse).

Access paths:

  • Cohere API β€” Chat endpoint, model ID command-a-plus-05-2026; free until rate limits on trial and production keys alike
  • Model Vault β€” Cohere's dedicated managed deployment, and the documented path to production beyond rate limits
  • Hugging Face β€” Apache 2.0 weights for self-hosting, plus a free trial Space
  • Oracle OCI β€” Cohere's multi-cloud docs use command-a-plus-05-2026 in their OCI example
  • Azure AI Foundry β€” available via Microsoft's model catalog
  • Self-hosted β€” 1Γ— B200 or 2Γ— H100 at W4A4

Because the weights are Apache 2.0, self-hosting carries no licensing fee.

Ecosystem & Tools

  • Cohere North: Cohere's enterprise AI platform, where Command models power agentic workflows over internal data.
  • Cohere Toolkit: Open-source templates for building RAG and agent applications on Command models.
  • Cohere SDKs: Official Python, TypeScript, Java, and Go clients for the Chat endpoint.
  • Hugging Face: Open weights, quantized variants, and a hosted demo Space.
  • Azure AI Foundry: Managed hosting for the W4A4 checkpoint.

Community & Resources

Frequently Asked Questions

Command A+ was released on May 20, 2026. Its API model ID is command-a-plus-05-2026.
No. Cohere has never shipped a model named Command 2.0, and has never used an N.0 version scheme for the Command family. The actual lineage is Command, Command R, Command R+, Command A, and Command A+. If you see a reference to Command 2.0, it is incorrect.
Yes. Command A+ is released under the Apache 2.0 license, and the weights are published on Hugging Face by Cohere Labs. The primary release is the W4A4 quantized checkpoint, CohereLabs/command-a-plus-05-2026-w4a4, with BF16 and FP8 variants also available.
218 billion total parameters with 25 billion active per token. It is Cohere's first Mixture-of-Experts model β€” every previous Command model was dense.
128K tokens of input context and up to 64K tokens of generation. Note that this is smaller than Command A (command-a-03-2025), which has a 256K context window but only 8K max output.
Not entirely. Cohere describes Command A+ as "the last model in the Command A family" and the fastest and most performant of the group, but command-a-03-2025 remains live and is still the right choice when you need a 256K context window.
Cohere publishes no per-token price. Instead, its docs state that "for both trial keys and production keys, Command A+ is free until rate limits are reached." Beyond those limits, production access runs through Cohere's Model Vault, whose hourly and monthly instance rates are on the pricing page. The Apache 2.0 weights can also be self-hosted at no license cost.
April 1, 2025, per Cohere's model documentation.
Cohere states that Command A+ fits on a single NVIDIA B200 at W4A4 quantization, or on two H100s at W4A4.
48 languages, including all official European Union languages. This is an expansion from the 23 languages covered by earlier Command A models.
Yes. It accepts text and image input and produces text, reasoning traces, and tool calls. Reasoning can be enabled or disabled per request, and a thinking token budget can be set. It is served through Cohere's Chat endpoint.

Related Models

Mistral Medium 3.5

Mistral AI's frontier-class multimodal model for agentic and coding work: a dense 128B model with a 256K context window, released April 2026 under a modified MIT license.

Llama 4

Meta's last open-weight frontier release: Scout and Maverick, natively multimodal Mixture-of-Experts models with 17B active parameters and up to a 10M token context window.

DeepSeek V4

DeepSeek's open-weight MoE flagship, shipped as V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B), with a hybrid sparse attention stack, a 1M token context window, and an MIT license.

Qwen3.7-Max

Alibaba Cloud's flagship Qwen model, announced May 20, 2026. A closed-weight, API-only agent model with a 1M token context window and hybrid thinking on by default β€” a break from the open-weight tradition that Qwen3.6 still carries.

GPT-5.5

OpenAI's flagship model for coding and professional work, with a 1,050,000 token context window, 128K max output, and reasoning effort levels from none through xhigh.

DBRX (Discontinued)

Databricks' open Mixture-of-Experts LLM, released March 2024 with 132B total / 36B active parameters. Fully retired from Databricks hosting by December 19, 2025, and its official Hugging Face weights are no longer published.

Explore More Models

Discover other AI models and compare their capabilities.

Browse Models← Back to Models