Command A+ Review: 218B Open MoE Model (2026)
Cohere Command A+: 218B-parameter open-weights MoE model, Apache 2.0, 128K context, 48 languages. Runs on 1x B200 or 2x H100. Released May 20, 2026.
Cohere Command A+ is an open-weights sparse mixture-of-experts model released May 20, 2026, with 218B total parameters (25B active), a 128K-token context window, and Apache 2.0 licensing. It unifies four prior Command A variants and supports 48 languages with major agentic benchmark gains over its predecessor.
Cohere Command A+, released May 20, 2026, is a 218B-parameter sparse MoE model (25B active) with Apache 2.0 open weights and a 128K-token context window. It unifies Command A, Command A Reasoning, Command A Vision, and Command A Translate, and runs on as few as 1x B200 or 2x H100 GPUs.
Provider: Cohere · Family: Command A
Context window: 128,000 tokens · Max output: 64,000
Input modalities: text, image, tool-calls · Output: text, tool-calls
About Cohere Command A+
Cohere Command A+ is a sparse mixture-of-experts foundation model released by Cohere on May 20, 2026 under the model ID command-a-plus-05-2026. It has 218 billion total parameters with 25 billion active per token across 128 experts (8 active per token), architected to run on far less hardware than its total parameter count suggests. Command A+ unifies capabilities that were previously split across four separate Cohere models, Command A, Command A Reasoning, Command A Vision, and Command A Translate, into a single scalable model, making it Cohere's first model to combine multimodal reasoning and machine translation in one checkpoint. It sits at the top of Cohere's Command lineup, above Command R+ and the smaller Command R7B. On benchmarks, Command A+ posts an Artificial Analysis Intelligence Index score of 37, a meaningful step up from earlier Command generations though still below frontier-tier models like GPT-5-series and Claude Opus. Cohere's own released comparisons versus its predecessor Command A Reasoning show large agentic gains: tau-squared-Bench Telecom jumped from 37% to 85%, and Terminal-Bench Hard (agentic coding) rose from 3% to 25%. Internal enterprise evals show a 20% improvement on agentic question answering and a 32% improvement on spreadsheet analysis tasks. On multimodal benchmarks, Command A+ scores 75.1% on MMMU (essentially matching Command A Vision's 75.3%) and improves MathVista from 73.5% to 80.6%. Independent tracker vals.ai lists a composite Vals Index accuracy of 24.69%, measured differently from Cohere's internal comparisons, and places Command A+'s context window at 128K tokens. Command A+ supports a 128,000-token input context window with a 64,000-token maximum generation length, unchanged in input size from Command A Reasoning but paired with meaningfully faster inference. Cohere reports output throughput up to 63% higher and time-to-first-token reduced 17% versus Command A Reasoning at full precision, with the W4A4 4-bit quantization adding a further 47% speed boost and 13% latency reduction, plus an additional 1.5-1.6x speedup available through speculative decoding. The model accepts text, image, and tool-use input and produces text and tool-use output, with genuine multimodal reasoning (not just image captioning) via its unified architecture. Language support expanded from Command A Reasoning's 23 languages to 48, covering all official EU languages, with tokenization efficiency gains of 20% for Arabic, 18% for Japanese, and 16% for Korean. Function calling and agentic tool use are core design targets, reflected in the large tau-squared-Bench and Terminal-Bench gains cited above. Command A+ is released under a full Apache 2.0 license with no usage restrictions, and weights are downloadable from Hugging Face (CohereLabs/command-a-plus-05-2026) in three quantization formats: BF16 (16-bit, requiring 4x B200 or 8x H100 GPUs), FP8 (8-bit, requiring 2x B200 or 4x H100), and W4A4 (4-bit, Cohere's recommended default for most deployments, requiring as little as 1x B200 or 2x H100). Cohere states benchmark quality differences across quantizations are negligible. There is no confirmed standard per-token hosted API price on Cohere's public pricing page as of this writing; Cohere's legacy Command R+ sits at $2.50/$10.00 per 1M tokens for comparison, but Command A+ is positioned primarily for self-hosted or Cohere Model Vault managed deployment rather than pay-as-you-go API billing. Deployment options include self-hosting the open weights directly, or using Cohere's Model Vault managed service, which handles infrastructure while keeping data inside a customer's own environment, a deployment model Cohere markets specifically at regulated industries needing sovereign or air-gapped AI. Oracle Cloud Infrastructure's Generative AI service has added Command A Reasoning and Command A Vision as managed offerings; Command A+'s cloud marketplace availability beyond direct Hugging Face and Model Vault access was not confirmed in available sources as of mid-2026. Command A+ carries Cohere's standard two-mode safety configuration: contextual mode for wide-ranging interactions with fewer output constraints while still rejecting clearly harmful or illegal content, and strict mode that avoids sensitive topics like violence, sexual content, and profanity entirely. Cohere has not published a dedicated third-party red-team partner list for Command A+ specifically, and detailed safety benchmark numbers (harmbench-style refusal rates) were not found in public sources at time of research; Cohere directs safety documentation requests to labs@cohere.com. Command A+ is best suited to enterprises that need self-hosted or sovereign-deployable AI (regulated industries, government, on-premises requirements) with multilingual and agentic coding capability, and teams that want a genuinely open Apache 2.0 model rather than a closed API-only frontier model. It is not the right choice for teams chasing the absolute highest reasoning benchmarks, where GPT-5-series and Claude Opus lead, or for teams without GPU infrastructure to self-host, since Command A+ has no confirmed low-cost hosted API tier the way OpenAI or Anthropic offer. Command A+ replaces four prior Command variants at once, so existing Command A Vision or Command A Translate users should expect a migration path onto the unified model going forward.
Pricing
Command A+ has no confirmed standard per-token hosted API price on Cohere's public pricing page as of mid-2026. It is open-weights (Apache 2.0) and positioned for self-hosting or Cohere Model Vault managed deployment. For reference, Cohere's Command R+ hosted API sits at $2.50 per 1M input / $10.00 per 1M output tokens; self-hosting Command A+ costs infrastructure only (minimum 1x B200 or 2x H100 at W4A4 quantization).
Key Features
- Unified Reasoning, Vision, and Translation: Combines four prior Command A variants into a single checkpoint, Cohere's first model to unify multimodal reasoning and machine translation.
- 218B/25B Active Parameter MoE: Sparse mixture-of-experts with 128 experts (8 active per token), enabling frontier-scale capacity on a fraction of the compute of a dense model that size.
- 48-Language Support: Expanded from Command A Reasoning's 23 languages, covering all official EU languages with measurable tokenization gains for Arabic, Japanese, and Korean.
- W4A4 4-bit Deployment: Runs on as few as 1x B200 or 2x H100 GPUs at 4-bit quantization, with Cohere reporting negligible quality loss versus full precision.
- Apache 2.0 Open Weights: Fully open license with no usage restrictions, downloadable from Hugging Face in BF16, FP8, and W4A4 formats.
Pros
- Full Apache 2.0 license with weights on Hugging Face, no usage restrictions or per-token lock-in.
- Major agentic gains over its predecessor: Terminal-Bench Hard 3% to 25%, tau-squared-Bench Telecom 37% to 85%.
- Runs on as few as 1x B200 or 2x H100 GPUs at W4A4 quantization, unusually accessible for a 218B-parameter model.
Cons
- Artificial Analysis Intelligence Index of 37 trails frontier closed models on composite reasoning.
- No confirmed low-cost hosted per-token API; effectively requires self-hosting or a Model Vault contract.
- 128K context window is smaller than several 2026 competitors offering 1M+ token windows.
Benchmarks
- mmmu: 75.1
- mathvista: 80.6
- artificial analysis intelligence index: 37
Frequently Asked Questions
What is Cohere Command A+ and who built it?
Cohere Command A+ is a sparse mixture-of-experts foundation model released by Cohere, a Toronto-founded AI company, on May 20, 2026 under the model ID command-a-plus-05-2026. It has 218 billion total parameters with 25 billion active per token across 128 experts, and it unifies capabilities previously split across four separate models: Command A, Command A Reasoning, Command A Vision, and Command A Translate. This makes it Cohere's first model to combine multimodal reasoning and machine translation in a single checkpoint. It sits at the top of Cohere's Command lineup, above Command R+ and Command R7B. On Cohere's own comparisons against its predecessor Command A Reasoning, Terminal-Bench Hard agentic coding rose from 3% to 25% and tau-squared-Bench Telecom from 37% to 85%. It carries an Artificial Analysis Intelligence Index of 37 and a 128K-token context window, and is released fully open under Apache 2.0.
How much does Cohere Command A+ cost per 1M tokens?
Command A+ has no confirmed standard public per-token hosted API price as of mid-2026; Cohere's public pricing page does not list it, unlike legacy models such as Command R+ ($2.50 input / $10.00 output per 1M tokens). Instead, Command A+ is released under Apache 2.0 with free, downloadable weights, so the real cost is infrastructure: a minimum of 1x B200 or 2x H100 GPU at W4A4 4-bit quantization, scaling up to 4x B200 or 8x H100 for full BF16 precision. On-demand cloud GPU rental for a 1x B200 W4A4 deployment runs roughly $55 for 24 hours, and 2x H100 roughly $72, though rates vary by cloud provider and region. Enterprises wanting a managed alternative to self-hosting can use Cohere Model Vault, a managed deployment service, but pricing for that requires contacting Cohere sales directly since no public per-token rate is listed.
What is Cohere Command A+'s context window and max output?
Command A+ supports a 128,000-token input context window with a 64,000-token maximum generation length, the same input size as its predecessor Command A Reasoning but with significantly faster inference. Cohere reports output throughput up to 63% higher and time-to-first-token reduced 17% versus Command A Reasoning, with W4A4 4-bit quantization adding a further 47% speed boost and 13% latency reduction, plus up to an additional 1.5-1.6x speedup from speculative decoding. This 128K window is smaller than several 2026 frontier competitors that offer 1M-token or larger context windows, such as Amazon Nova 2 Pro or Gemini 3.1 Pro, making Command A+ better suited to standard document and conversation lengths than to extreme long-context retrieval tasks.
How does Cohere Command A+ compare on benchmarks vs GPT-5 and Claude?
Command A+ scores an Artificial Analysis Intelligence Index of 37, notably below frontier closed models like GPT-5-series and Claude Opus 4.8, which score in the 60s on the same composite index. On vals.ai's independent tracking, Command A+ posts a Vals Index accuracy of 24.69%. Where Command A+ shows its largest gains is against its own predecessor rather than external rivals: Terminal-Bench Hard agentic coding rose from 3% to 25% and tau-squared-Bench Telecom from 37% to 85% compared to Command A Reasoning. On multimodal evals, Command A+ scores 75.1% on MMMU and 80.6% on MathVista. In practice, this benchmark gap versus frontier closed models means Command A+ is not the right choice for teams needing the absolute highest reasoning accuracy; its strength is being a genuinely open, self-hostable model with strong agentic and multilingual gains rather than a frontier-reasoning leader.
Is Cohere Command A+ open source or proprietary?
Command A+ is fully open-weights under the Apache 2.0 license, one of the most permissive open-source licenses available, with no restrictions on commercial use. Weights are downloadable from Hugging Face under CohereLabs/command-a-plus-05-2026 in three quantization formats: BF16 (16-bit, requiring 4x B200 or 8x H100 GPUs), FP8 (8-bit, requiring 2x B200 or 4x H100), and W4A4 (4-bit, Cohere's recommended default, requiring as little as 1x B200 or 2x H100). Cohere states benchmark quality differences across these quantizations are negligible. For teams that don't want to self-host, Cohere also offers Model Vault, a managed deployment service that keeps data inside the customer's environment rather than running on shared multi-tenant infrastructure. There is no separate closed or gated version of Command A+; the open release is the primary distribution channel.
What modalities does Cohere Command A+ support?
Command A+ accepts text, image, and tool-use input, and generates text and tool-use output, making it a genuine multimodal reasoning model rather than a text-only model with a bolted-on vision adapter. It does not support audio or video input or output. Function calling and agentic tool use are core strengths, reflected in large benchmark gains on tau-squared-Bench Telecom (37% to 85%) and Terminal-Bench Hard (3% to 25%) versus its predecessor. On multimodal image reasoning, it scores 75.1% on MMMU, roughly matching the prior specialized Command A Vision model's 75.3%, and improves MathVista performance from 73.5% to 80.6%. Compared to competitors like Amazon Nova 2 Pro (which adds video input) or GPT-5-series and Gemini (which add native audio), Command A+'s modality set is narrower but its unification of reasoning, vision, and translation into one model is a structural differentiator Cohere had not offered before.
Does Cohere Command A+ train on user data?
For self-hosted deployments, there is no data-training concern since Command A+ runs entirely on the customer's own infrastructure with weights downloaded once from Hugging Face; the model does not phone home or transmit usage data by default. For customers using Cohere Model Vault, the managed deployment option, Cohere's positioning is that data stays inside the customer's own environment rather than shared multi-tenant infrastructure, which is the core selling point for regulated-industry and sovereignty-focused buyers. Cohere has not published Command A+-specific SOC 2, ISO 27001, or HIPAA certification details separate from its general enterprise compliance program, and specific compliance certifications were not confirmed in available public sources at time of research; enterprise buyers evaluating Model Vault for regulated workloads should request Cohere's current compliance documentation directly.
Who is Cohere Command A+ best for and who should avoid it?
Command A+ is best for regulated enterprises and government buyers needing sovereign or on-premises AI deployment, teams building multilingual agentic applications across its 48 supported languages, and organizations that specifically want a genuinely open Apache 2.0 model rather than a closed API-only frontier model, since it can be audited, fine-tuned, and run air-gapped. It's also a strong fit for teams with existing GPU infrastructure who prefer self-hosting economics over recurring per-token API billing. Teams chasing the absolute highest reasoning benchmark scores should avoid it and look at GPT-5-series or Claude Opus instead, both of which score well ahead on the Artificial Analysis Intelligence Index. Solo developers or small teams without GPU access should also avoid it, since there's no confirmed low-cost hosted per-token API tier the way OpenAI or Anthropic offer; the practical entry cost is a minimum of 1x B200 or 2x H100 GPU. Teams needing context windows above 128K tokens for extreme long-document analysis should also look elsewhere.