Jamba 1.7

What it is

Jamba 1.7 is an update to the Jamba family, published on Hugging Face in July 2025; the related collection was updated on 2 July 2025.^[1] It comes in two sizes. Jamba Large 1.7 has 398B total parameters and 94B active parameters, while Jamba Mini 1.7 has 52B parameters.^[1][2] The distinction between total and active parameters is useful for reading the specifications: total parameters describe the full model, while active parameters describe the part engaged in a given computation. Jamba 1.7 supports a 256k-token context and the same nine languages as earlier versions.^[1] A context window is the amount of text a model can consider at once, so a 256k-token limit matters for long documents, long conversations, or prompts that must keep several pieces of material in view. The nine-language coverage means the family is aimed at multilingual use rather than English only.

What changed

The update reports better grounding and steerability than version 1.6. On grounding, the FACTS score rose from 0.758 to 0.832 for Jamba Large 1.7 and from 0.727 to 0.790 for Jamba Mini 1.7.^[1][2] Grounding refers to how well a model keeps its answers tied to the documents it is given; FACTS is the measure reported here. On AI21's instruction-following benchmark, the steerability score rose from 0.782 to 0.84 for Large and from 0.68 to 0.76 for Mini.^[1][2] Steerability refers to how well a model follows directions and constraints. The figures place 1.7 against 1.6, so the update is presented as an improvement in factual grounding and instruction following.

Both model cards list a knowledge cutoff of August 22, 2024.^[1][2] A knowledge cutoff marks the date up to which built-in knowledge is documented; events after that date should not be assumed to be part of the model's knowledge.

Architecture and license

Jamba Mini 1.7's model card says it builds on the SSM-Transformer hybrid architecture and efficiency gains of previous versions.^[2] An SSM-Transformer hybrid combines elements of state space models and Transformer models; the model card describes Mini 1.7 as continuing that design line. Jamba 1.7 is released under the Jamba Open Model License, which allows research and commercial use under its terms.^[1] That matters because researchers and organizations can use the model for research and commercial work within the license terms.

Where it sits in the Jamba line

Jamba 1.7 is the Jamba update reported in July 2025, keeping the 256k-token context and nine languages while reporting higher FACTS and instruction-following scores than 1.6.^[1][2] Large offers 398B total parameters and 94B active parameters; Mini has 52B parameters and continues the hybrid architecture and efficiency gains of earlier versions.^[1][2] The two versions share the same license and the same knowledge cutoff of August 22, 2024.^[1][2] Jamba 1.7 is therefore a family of two sizes under one open model license, documented around long context, multilingual coverage, grounding, and instruction following.

^[1]: AI21-Jamba-Large-1.7 model card, Hugging Face (AI21 Labs). ^[2]: AI21-Jamba-Mini-1.7 model card, Hugging Face (AI21 Labs).