Jamba 1.5 Mini was released on the same day as its larger sibling, 22 August 2024, and carries the same architecture in a form small enough to run in ordinary places: Mamba state-space layers interleaved with attention, a sparse mixture-of-experts, 52 billion parameters with 12 billion active per token. Its significance is the ratio. In the RULER long-context benchmark AI21 published with the release, Mini held its full claimed 256,000-token window with a score of 86.1 at the longest setting — better at 128K than GPT-4 Turbo, Llama 3.1 70B or Mistral Large 2, all of which are several times its size and none of which reached their advertised limits. On general knowledge it sits where its size suggests, with 69.7 on MMLU and 46.1 on Arena Hard against the Large model's 81.2 and 65.4, which is the honest trade: the hybrid buys context length, not raw reasoning. AI21 shipped it with open weights under the Jamba Open Model License and into every major cloud marketplace. The generation has since been superseded twice over — by Jamba 1.6 and 1.7, then by Jamba 2 — and left AI21's own API on 6 May 2025, but Amazon Bedrock still serves it until 26 November 2026.