Cohere Command A+ released as open-source under Apache 2.0

Source: The Decoder (Maximilian Schreiner), 2026-05-21. Link: The Decoder coverage

TL;DR

Cohere released Command A+ as open-source weights under Apache 2.0. It is the company's most powerful language model to date. The Decoder coverage is short on technical detail, so this page is a placeholder pending Cohere's full model card and benchmark disclosures.

Why this matters

Command A+ joins a growing list of frontier-class open-weight models released in May 2026, including the open-model wave through May 21 documented in the 2026-05-17 page and Tencent's Hy-MT2 translation family (1.8B / 7B / 30B-A3B) released the same day on r/LocalLLaMA. Cohere's Apache 2.0 license is the most permissive of the May-2026 cohort. The release reinforces the AI Snake Oil position that frontier capability is "at most a few months" away from open weights, which is load-bearing for the Kapoor argument that nonproliferation alone is brittle.

Cross-references

Open questions

Benchmark numbers vs Mythos, GPT-5.5, Gemini 3.5.
Parameter count and training-token disclosure.
Whether Cohere has retained any commercial-use restrictions in addendum to Apache 2.0 (e.g., trademark or attribution).

Source

Raw: raw/rss/2026-05-21-the-decoder-cohere-open-sources-its-strongest-model-yet.md.