AI21 Labs releases Jamba 1.6 and launches Maestro

Background

AI21 Labs' Jamba line is aimed at organisations that want to run a model inside their own infrastructure rather than call a hosted endpoint. That requirement shapes a release: the weights have to be obtainable, the licence has to be legible to a legal team, and the hardware footprint has to be manageable. Jamba 1.6 is organised around those points.

What was released

In March 2025 AI21 Labs released Jamba 1.6, packaged for private enterprise deployment^[1]. The release covers at least two sizes, Large and Mini, and the weights were published under the Jamba Open Model License^[2]. Putting weights out under a named licence, rather than only offering an API, is what makes private deployment concrete: an organisation can take the model files and run them in its own environment, subject to that licence's terms.

Size and efficiency

Jamba Large 1.6 has 398B total parameters, of which 94B are active^[2]. The gap is worth explaining. When a model reports fewer active parameters than total parameters, only part of the network is used for each token, so the compute spent per token tracks the active count more closely than the total. Total parameters correspond roughly to capacity; active parameters correspond to what one forward pass costs. For a buyer, the second number is the one that decides how much hardware to provision.

Benchmark results

AI21 reported that Jamba Large 1.6 beat Mistral Large 2, Llama 3.3 70B and Command R+ on Arena Hard, CRAG and LongBench^[2]. The three probe different things: Arena Hard compares models on hard prompts, CRAG tests question answering that depends on retrieved external material, and LongBench tests long-document understanding. Read together, the claim is that the model holds up on difficult requests, on tasks that require looking things up, and on long inputs. The model card for the smaller Jamba Mini 1.6 reports Arena Hard 51.2 and CRAG 76.2^[3].

Maestro

The same month AI21 launched Maestro, a planning and orchestration system^[1]. Orchestration means not handing a whole task to a single model call: the system breaks a goal into steps, decides which model or tool handles each step, and assembles the results. It is complementary to open weights — one addresses how several models are organised to do work, the other whether a model can live in your own data centre.

Why it matters

For an enterprise buyer, a single release containing downloadable weights, a named licence, published benchmark numbers and an orchestration layer removes several rounds of stitching: where the model comes from, what it costs to run, roughly how well it performs, and how multiple models are coordinated are all answered in one place. How far those numbers carry over to a particular business is something each buyer has to test.