AI21 Labs releases Jamba 1.7

Jamba Large 1.7 and Jamba Mini 1.7 are updated versions of AI21 Labs' Jamba family of language models, published on Hugging Face in July 2025^[1]. The release was an update to an existing line rather than a new design: the change the model cards emphasise most is a push to make the models answer strictly from the material placed in front of them, and to follow instructions more reliably than the previous version^[1][2].

Background

AI21 Labs publishes Jamba in two sizes, Large and Mini, and numbers the releases, so that 1.7 succeeds a version 1.6^[1][2]. Both 1.7 models were posted on Hugging Face, the model-sharing site where AI21 Labs hosts the model cards that document their behaviour and specifications^[1].

What the update targets

The stated focus of the release is grounding: getting the model to base its answer on the context it is given, such as a retrieved document, a contract or a manual passed in with the question, instead of on whatever it absorbed during pretraining^[1]. This is the failure mode that matters most when a model is placed inside a document workflow, because an answer that drifts away from the supplied text is hard for a reader to check and hard for a system to trust.

The model cards quantify the change with a FACTS grounding score, a measure of how well answers stay tied to the provided context. For Jamba Large 1.7 that score rose to 0.832 from 0.758 in the previous version^[1]. Instruction following also improved over version 1.6, according to both model cards^[1][2].

Specifications

Two things were kept unchanged from the earlier release: the models accept a 256k-token context, and they are distributed under the Jamba Open Model License^[1]. A context window of that size is the amount of text the model can take into account at once, which is what allows a long document to be supplied in full rather than summarised first.

The Jamba Large 1.7 model card lists 94B active and 398B total parameters^[1]. The two figures describe different things: the total count covers all the weights in the network, while the active count covers only the part engaged for each token as the model runs. A card that reports both is describing a design in which only a fraction of the network is used at any one step, so compute per token is closer to the active figure than to the total.

The model card also lists a knowledge cutoff date of August 22, 2024^[1]. The cutoff marks the end of the material the model was trained on, and everything after it has to reach the model through the context window instead, which is part of why grounding behaviour and a long context are discussed together.

What it means

For users of the Jamba line, 1.7 is a drop-in update: same context length, same licence, better scores on the two behaviours that decide whether a model is usable over long supplied documents^[1][2]. The grounding number is the more informative of the two, because it turns a vague complaint, that the model made things up, into a figure that can be compared across versions.