AI Potluck
Model components / Fine-tuned / chat models

Command R+

Cohere

Cohere's enterprise foundation model line, currently anchored by Command A+ (released May 20, 2026), a 218B-parameter sparse MoE (25B active) with native citation generation, W4A4 lossless quantization, and the company's first fully Apache 2.0 licensed frontier model. Runs on a single B200 or two H100s. Positioned for sovereign AI and enterprise RAG with native support for 48 languages and explicit grounding spans linking every factual claim to source documents. Earlier Command R+ (104B, 2024) is the predecessor; with Command A+ the line has flipped from API-only to open-weights.

Correctly an instruct model, belongs in finetuned_chat (was miscategorized as base elsewhere). Openness and capability were both rescored onto the governing release, Command A+ (Apache 2.0, ungated), on 2026-08-14; all three axes now read the same release the description and the declared artifacts follow. Verified 2026-08-14 via the Command A+ model card and the CohereLabs org listing.

Openness

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

Openness is read on the line's current release, Command A+ (command-a-plus-05-2026), which the product declares as an artifact. Its card reads "Command A+ is an open source model with 25 billion active parameters and 218B total parameters" under "License - Apache 2.0", and the repository is tagged Apache-2.0 and ungated. The CohereLabs org listing confirms it is the newest Command release and that all three of its SKUs - bf16, fp8 and w4a4 - carry Apache-2.0, so resolving to the most restrictive license across the release's SKUs still lands on Apache-2.0. That is an OSI license, which takes the ladder off the restricted rung. The superseded 2024 104B checkpoint c4ai-command-r-plus is still CC-BY-NC-4.0 and auto-gated, but it does not govern. It does not reach 4 or 5 either: the post-training mixture and the fine-tuning pipeline are both unpublished, so the score settles at the open-weights center of gravity. Cohere's acceptable use policy is not listed as a license because it does not attach to Apache-2.0 weights.

  • https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16 recorded 2026-08-14

    the governing release - "Command A+ is an open source model with 25 billion active parameters and 218B total parameters ... optimized for agentic, multilingual, and reasoning-heavy tasks with a focus on enterprise performance"; "License - Apache 2.0" and `"gated":false`; BF16, FP8 and W4A4 quantizations offered; no post-training mixture and no fine-tuning pipeline published

  • https://huggingface.co/api/models?author=CohereLabs&sort=createdAt&direction=-1&limit=40 recorded 2026-08-14

    CohereLabs newest-first. The Command line's newest release is command-a-plus-05-2026, shipped 2026-05-11 (bf16), 2026-05-18 (fp8, w4a4), all three tagged `license:apache-2.0`. Nothing newer in the Command line exists; the 2026-06/07/08 repos are the separate North line and the transcribe models. Every pre-2026-03 Cohere checkpoint, including c4ai-command-r-plus, is `license:cc-by-nc-4.0`.

  • https://huggingface.co/CohereLabs/c4ai-command-r-plus recorded 2026-08-13

    `license:cc-by-nc-4.0` in the repo tags with `"gated":"auto"`; 104B instruct model; the 2024 release, superseded and no longer governing

Adoption

3 high confidence
3.0

A tier bands on its strongest release, and this line is anchored by Command A+. The declared artifacts are command-a-plus-05-2026-bf16 (84,821) and c4ai-command-r-v01 (24,775), summing to 109,596 Hugging Face downloads in the trailing 30 days, which bands at 100K-1M, level 3. Deliberately excluded: command-a-vision-07-2025, the transcribe models and the embed models, which are separate Cohere product lines rather than SKUs of this one, and counting them would over-attribute. The superseded 2024 Command R+ checkpoint reads about 3,258 a month on its own, gated and held back by a non-commercial license and its 104B size - hobbyist and legacy scale, and not what the tier bands on.

Capability

3 medium confidence
3.0

A tier is scored on its strongest release, which here is Command A+ (05-2026), a 218B-total, 25B-active mixture-of-experts model Cohere positions for "agentic, multilingual, and reasoning-heavy tasks with a focus on enterprise performance", with vision inputs, a 128K context and trained-in conversational tool use. The score rests on that feature set rather than on a benchmark, and that is the honest part of the placement: the Command A+ card publishes no benchmark table at all. 3 rather than 4 for that reason, and because the 2026 reasoning frontier in this category sits above it on published scores - Nemotron at 5, DeepSeek-R1 and Hermes at 4, all on AIME, GPQA and SWE-bench evidence Cohere does not publish. It lands level with the other capable but unbenchmarked general instruct models here, Llama Instruct and Qwen Instruct, both at 3. The superseded 2024 Command R+ 104B checkpoint and its legacy Open LLM Leaderboard breakdown (average 74.6, MMLU 75.7) would be a fair 2 against a 2026 comparison set, but that is not the release the tier scores on. No peer at this level has been checked recently enough to serve as a comparison anchor, so none is recorded.

  • https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16 recorded 2026-08-14

    the governing release - 25B active / 218B total, "optimized for agentic, multilingual, and reasoning-heavy tasks with a focus on enterprise performance", vision inputs, 128K context, and a tool-use section stating the model "has been specifically trained with conversational tool use capabilities". No evaluation or benchmark table anywhere on the card.

  • https://huggingface.co/CohereLabs/c4ai-command-r-plus recorded 2026-08-13

    the 2024 checkpoint the old score read, with its legacy Open LLM Leaderboard breakdown; superseded

Verified 2026-08-13