abab 6.5

What it is

abab 6.5 is a family of large language models released by the Chinese AI company MiniMax on 17 April 2024, built on a mixture-of-experts (MoE) architecture^[1]. It belongs to the company's abab line of text models, which MiniMax kept proprietary rather than releasing as open weights; in January 2025 the company instead brought out the open-weight MiniMax-01 series^[1].

How the abab line got there

MiniMax began research and development in January 2022, a month after the company was founded^[1]. Its first text model, abab1, launched in April 2022; abab2 finished training in June 2022; abab3, with more parameters than its predecessors, followed in October 2022^[1]. Over the same period the company also shipped consumer apps: Glow in October 2022, Talkie in June 2023 and Xingye in September 2023^[1].

The model immediately before abab 6.5 was abab 6, launched in January 2024 and described by Pandaily as China's first mixture-of-experts large language model^[2]. Pandaily reported that MiniMax followed abab 6 with abab 6.5 in April 2024^[2].

What mixture-of-experts means

A conventional "dense" language model applies all of its parameters to every token it processes, so making the model larger makes every step more expensive. A mixture-of-experts model splits its feed-forward layers into many separate "expert" subnetworks and adds a routing step that sends each token to only a small number of them. The total number of parameters can therefore grow much faster than the computation used per token, at the cost of a more complicated training and serving setup: memory has to hold all the experts even though only some are used at any moment, and the router itself has to learn a workable division of labour. Announcing abab 6.5 as an MoE family places it in this design tradition, and its predecessor abab 6 had already been presented as an MoE model^[2].

Long context and the needle test

Pandaily reported that abab 6.5 supports a context length of 200,000 tokens^[2]. The context length is the amount of text a model can keep in view at once — the input plus whatever it has generated so far — and it bounds tasks such as summarising a long report, searching a code base or holding a long conversation. Tokens are the chunks into which text is split before the model reads it, so a window measured in tokens is a measure of how much raw text the model can attend to in a single pass.

Pandaily also reported that abab 6.5 answered 891 of 891 queries correctly in a needle-in-a-haystack test^[2]. In that kind of test a short, arbitrary sentence is hidden at some position in a long document and the model is asked to retrieve it; moving the sentence around and lengthening the document probes whether the model's attention actually reaches every part of a long input rather than only its beginning or end. A perfect score across 891 queries indicates reliable retrieval within the tested range.

Retrieval, though, is only one part of long-context ability. Finding a planted sentence is easier than reasoning over a whole document — comparing sections, following an argument, or noticing what is missing — and a needle test does not by itself establish those skills.

Where it sits in the line

Within MiniMax's own history, abab 6.5 is one step in a sequence that runs from abab1 in April 2022 through abab2, abab3 and abab 6, and on to the open-weight MiniMax-01 series in January 2025^[1]. The same company was building consumer products at the same time — Glow, Talkie and Xingye — so the text models and the apps developed alongside each other^[1].