Command A Plus

Introducing Command A+ | Cohere North Mini Code. Cohere's first model for developers. Learn more

May 20, 2026

4 minute read

Introducing Command A+: Making sovereign agentic capabilities available to all Our fastest and most powerful language model yet. Command A+ is an open-source enterprise workhorse built for complex reasoning, multimodal and multilingual agentic tasks — all while running on as little as two H100 GPUs.

Today, we’re releasing Command A+ open-source. A mixture-of-experts (MoE) model, Command A+ is an efficient, versatile, and privately deployable LLM built for high-performance agentic tasks with minimal compute overhead.

Born from a year of deploying North with our customers, it surpasses every previous generation in the Command series and unifies their capabilities into a single scalable model.

Now freely available under an Apache 2.0 license , Command A+ advances Cohere’s mission to make sovereign AI a technological reality — giving developers direct access to enterprise-grade agentic capabilities across experimentation, deployment, and production workflows.

Visit Hugging Face to download the weights - available in several near lossless quantizations - and read our implementation guides. For a dedicated, managed inference environment, deploy Command A+ in Model Vault today. Snapshot

Model command-a-plus-05-2026

License Apache 2.0

Architecture Sparse / MoE

Model size 218B total; 25B active

Context length 128K input context; 64K max generation

Input modalities Text, image, tool use

Output modalities Text, reasoning, tool use

Languages Supports 48 languages. Full list

Optimized for Reasoning, agentic workflows, RAG, multilingual, multimodal document processing

Supported frameworks vLLM, Transformers

Hardware (minimum) 1× B200 @ W4A4 2× H100s @ W4A4

Northwards For the past year, North — Cohere’s integrated enterprise workspace for building and deploying agentic AI — has been the driving force behind much of our innovation. Through that work, we set out to build a unified model for customers that simplifies deployment, can run locally, and synthesizes capabilities from across the Command family.

The work is already paying off. Read how our customers have been using North to transform their operations .

However, sovereign AI is much bigger than Cohere. Empowering engineers with models that they can run, control, and adapt themselves is the most acute challenge facing this generation of AI.

We’ve optimized Command A+ for practical, developer-focused use, including support for low-bit quantization, efficient inference, and integration across open inference frameworks. AI independence for all.

We can’t wait to see what the community builds. Command, consolidated Command A+ outperforms previous Command A models in key dimensions of enterprise workloads, including multimodal understanding, retrieval, long-horizon, and complex reasoning.

Command A+

Command A

Command A Reasoning

Command A Vision

Command A Translate

Size

218B A25B

111B 111B 112B 111B

Reasoning

✓

—

✓

— —

Multimodal

✓

— —

✓

—

Tool use

✓

✓

✓

— —

Multilingual

48

23 23 6 23

Image 2: comparing the capabilities of Command A+ with other models in the Command A family.

Compared with Command A Reasoning, 𝜏²-Bench Telecom scores improved from 37% to 85%, with agentic coding performance on Terminal-Bench Hard reaching 25% from 3%. Gains were also achieved on non-agentic reasoning, instruction following, and other code generation tasks. Image 3: Performance for Command A+ and Command A Reasoning on a range of popular open-source benchmarks. See footnote for further details. 1

Command A+ performs strongly within North applications, reflecting its original design goals. Agentic Question Answering accuracy and spreadsheet analysis quality improved by 20% and 32% over Command A Reasoning, respectively. Memory performance — testing North’s skill in reasoning across conversations and stored data — scored 54% with Command A+ compared to 39% with Command A Reasoning. Image 4: Performance improvements on three internal evaluations on North. Agentic Question Answering measures how well a model can answer enterprise questions using MCP-connected cloud file systems. Data Analysis scores a model’s ability to perform data science tasks over uploaded spreadsheets, and Memory Usage Quality measures how well an agent can leverage information in North’s memory system from a previous session to answer questions in a subsequent session. All are scored using LLM-as-a-judge techniques. For multimodal understanding and reasoning, Command A+ achieved 63% on MMMU Pro and 75.1% on MMMU, (compared with 65.3% for Command A Vision for the latter). MathVista scores increased from 73.5% to 80.6%, and CharXiv reasoning improved from 46.9% to 52.7%, reflecting broad gains across document understanding tasks. Image 5: Comparison of multimodal performance for Command A+ and Command A Vision. Command A+ is Cohere’s first multimodal reasoning model and provides significant boosts (compared with Command A Vision) to relevant tasksets such as CharXiv reasoning. We follow standard methodology for the given benchmarks. Command A+ significantly expands multilingual capability, broadening language coverage from 23 to 48 languages and recording gains in machine translation and multilingual reasoning. Image 6: comparison of multilingual performance for Command A+ and Command A Reasoning. MT-AIME 2025 is an internal translation of AIME-2025 — an English-language mathematics benchmark — evaluated for Arabic, Japanese, and Korean. WMT24++ is a widely used public benchmark, evaluated here for xCOMETxl. 2

Command A+ achieved a score of 37 on the Artificial Analysis Intelligence Index , outperforming other leading open models, reflecting its strength as a general-purpose model for enterprise agentic workflows [3]. Efficiency at scale Efficiency is a core constraint in enterprise AI deployment. It determines whether a language model can be deployed practically at scale by shaping the compute, memory, latency, power, and infrastructure required to serve it reliably and cost-effectively.

We engineered Command A+ to be extremely hardware efficient. The model is available today on Hugging Face in 16-bit (BF16), 8-bit (FP8), and 4-bit (W4A4) quantizations, with imperceptible differences in quality. In practice, this enables Command A+ to run on as little as two NVIDIA H100s or a single NVIDIA Blackwell GPU, with...