Jamba 1.5 (Mini and Large)

What it is

Jamba 1.5 is a pair of instruction-tuned language models from AI21 Labs, released in two sizes called Mini and Large, that combine Transformer and Mamba layers inside a single architecture. They are described in a paper posted to arXiv on 22 August 2024 ^[1]. The two models differ chiefly in scale: Jamba 1.5 Large has 398 billion parameters in total, of which 94 billion are active, while Jamba 1.5 Mini has 52 billion in total and 12 billion active ^[2]. The gap between the total and active counts means that only part of the network is engaged when a given piece of text is processed, so the cost of each step stays below what the full parameter count would suggest.

Why mix two architectures

A Transformer layer uses attention, in which each token is compared with the tokens around it. That is expressive, but the amount of computation grows quickly as the context gets longer. Mamba belongs to the family of state-space models, which read a sequence step by step while carrying a compact internal state, so the cost of processing grows roughly in proportion to length. Jamba 1.5 places both kinds of layers in one model, so that long inputs do not have to rely entirely on the part whose cost rises fastest with length. The models are also instruction-tuned, meaning they are trained to carry out user requests rather than merely to continue a piece of text.

Long context and serving efficiency

Both models have an effective context length of 256,000 tokens ^[1]. The paper describes that as the largest effective context among open-weight models ^[1]. A long context is only useful if it can actually be served, and here AI21 introduced a quantisation method called ExpertsInt8, which lets Jamba 1.5 Large serve 256k-token contexts on a single machine with eight 80 GB GPUs without any loss of quality ^[1]. Quantisation means storing weights in a lower-precision numerical format so that they take up less memory; the point of the method here is to make very long inputs affordable on ordinary hardware rather than in a large cluster.

What it can do

The models support English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic and Hebrew ^[2]. They also offer function calling, JSON mode and grounded generation, the last meaning answers produced with reference to material supplied to the model ^[2]. Jamba 1.5 was trained to produce valid JSON when asked to ^[2], and it supports tool use, so a model can call external functions instead of answering only from its own weights ^[2]. Those features matter for applications in which a model has to sit inside a larger piece of software and return results that a program can parse.

Benchmark results

The model card reports the following pairs of scores, Mini first and Large second: Arena Hard 46.1 / 65.4; MMLU with chain-of-thought prompting 69.7 / 81.2; MMLU-Pro 42.5 / 53.5; GSM8K 75.8 / 87; and BFCL 80.6 / 85.5 ^[2]. Arena Hard is a set of difficult prompts scored by comparison against other models, MMLU and MMLU-Pro are multiple-choice knowledge tests of increasing difficulty, GSM8K is a set of school-level arithmetic word problems, and BFCL is a function-calling benchmark. The pattern is consistent across all five: the larger model scores higher, and the gap is widest on Arena Hard, the comparison-style evaluation.

Release and what followed

The weights were released under the Jamba Open Model License, which permits research and commercial use under its terms ^[1][2]. In September 2024, AI21 and AWS announced a partnership under which the models are offered on Amazon Bedrock ^[3], which put them in front of customers who build on that cloud platform rather than downloading weights. The family was followed by Jamba 1.6 in March 2025 ^[3].

Jamba 1.5 is best understood as an attempt to get two things at once: the quality of a Transformer-based model and the cheaper long-sequence processing of a state-space design, with serving work done so that the advertised context length is usable on a single eight-GPU machine.