AI Potluck
Contribute
Back to Gap Map Model components / Fine-tuned / chat models

Command R+

Cohere
open weights / Overall score: 3.0

Command A+ is Cohere's enterprise language model, a 218B-parameter sparse mixture-of-experts (25B active) with vision input, conversational tool use and native citation generation through grounding spans that link claims to source documents. It supports 48 languages and a 128K context window, and runs on a single B200 or two H100 GPUs with W4A4 quantization. It succeeds the 104B Command R+ of 2024.

Scored against Command A+ rather than the earlier Command R and Command R+ checkpoints.

Openness

3 high confidence
3.0
weights
open(ungated, Apache-2.0)
data
closed
code
closed
license
Apache-2.0(OSI)

Command A+ ships under Apache-2.0 in all three quantization variants with no access gate. Its training data and fine-tuning pipeline are not published, which makes it an open-weights release rather than a fully open model; the earlier 104B Command R+ checkpoint remains under a non-commercial license.

  • https://cohere.com/blog/command-a-plus recorded 2026-09-29

    Cohere launch post dated May 20, 2026: "Command A+ outperforms previous Command A models in key dimensions of enterprise workloads"; lists a sparse MoE with 218B total and 25B active parameters, 128K input context, 48 languages, Apache 2.0 and 1x B200 or 2x H100 at W4A4.

  • https://huggingface.co/api/models?author=CohereLabs&sort=createdAt&direction=-1&limit=40 recorded 2026-09-29

    CohereLabs newest-first: the newest Command release is "CohereLabs/command-a-plus-05-2026-w4a4" together with the fp8 and bf16 repositories, each tagged "license:apache-2.0"; later repositories belong to the separate North, transcribe and tiny-aya lines, and no newer Command model exists.

  • https://huggingface.co/api/models/CohereLabs/command-a-plus-05-2026-bf16 recorded 2026-09-29

    Hugging Face API record for the bf16 repository: the tag "license:apache-2.0" is present and the record is not gated (gated false), so the weights download without an access gate; the repo id is "CohereLabs/command-a-plus-05-2026-bf16".

  • https://huggingface.co/CohereLabs/c4ai-command-r-plus recorded 2026-09-29

    Earlier 104B Command R+ checkpoint, superseded: repository tag "license:cc-by-nc-4.0" and a notice "You need to agree to share your contact information to access this model".

  • https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16 recorded 2026-09-29

    Model card for the governing release: "Command A+ is an open source model with 25 billion active parameters and 218B total parameters" and "has been specifically trained with conversational tool use capabilities"; "Apache 2.0" is the stated license. Searched the body for training data, post-training and fine-tuning pipeline publication: none; no evaluation table.

Adoption

3 high confidence
3.0

Adoption is measured on the current release, Command A+, using downloads from its two declared Hugging Face repositories; Cohere's separate vision, transcription and embedding lines are excluded. The earlier 104B Command R+ checkpoint sees far fewer downloads and remains gated under a non-commercial license, consistent with legacy use.

Capability

3 medium confidence
3.0

Command A+ is a 218B-parameter mixture-of-experts model (25B active) with a 128K context window, vision inputs, conversational tool use, 48 languages, and citation and grounding spans. Cohere's launch post compares it only with earlier Command models and cites no independent leaderboard result, so it sits with the Llama Instruct and Qwen Instruct models rather than above them.

  • https://cohere.com/blog/command-a-plus recorded 2026-09-29

    Launch post gives benchmark charts against earlier Cohere models only, plus "63% on MMMU Pro and 75.1% on MMMU" and "MathVista scores increased from 73.5% to 80.6%"; no independent leaderboard comparison is cited.

  • https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16 recorded 2026-09-29

    Grounded generation: "one can ask the model to include grounding spans (citations) in its response"; the card lists 25B active and 218B total parameters and W4A4 running on "1 x B200".

Verified 2026-09-27