MiniMax launches the abab 6.5 mixture-of-experts series

Background

MiniMax started research and development in January 2022, a month after the company was founded, and it moved quickly: its first text model, abab1, launched in April 2022^[1]. abab2 finished training in June 2022, and abab3, which carried more parameters, followed in October 2022^[1]. Every model in the line kept the abab prefix.

The line changed shape at the start of 2024. In January 2024 MiniMax launched abab 6, which Pandaily described as China's first mixture-of-experts large language model^[3]. In March 2024, a month before abab 6.5 appeared, Alibaba led a US$600 million round in MiniMax at a US$2.5 billion valuation^[1][2].

What was released

On 17 April 2024 the company launched the abab 6.5 series, a family of mixture-of-experts language models^[1]. Pandaily reported that it followed abab 6^[3].

Mixture of experts is an architectural idea rather than a single product. Instead of running every parameter of the network for every token, the model is built out of many parallel sub-networks, usually called experts, plus a routing component that decides which few of them each token should be sent to. Because a token only passes through a small part of the network, the computation spent per token stays below what a dense model of the same total size would need. That is the trade the architecture offers: builders can raise the total parameter count without raising per-token cost in the same proportion. The catch is that routing has to work well, or capacity sits unused.

Key details

Two figures reported by Pandaily describe what the release was meant to do.

The first is context length. Pandaily reported that abab 6.5 supports a 200,000-token context length^[3]. Context length is the amount of text a model can take in at once — the working material for a single request. A longer window means more of a document, a conversation or a code base can be held in view together instead of being cut into pieces and fed in separately.

The second is a retrieval test. Pandaily reported that abab 6.5 answered 891 of 891 queries correctly in a needle-in-a-haystack long-context test^[3]. The test works by hiding one short piece of information somewhere inside a very long document and then asking the model to find it. It is a narrow measure — it checks whether the model can still locate a detail, not whether it can reason over the whole document — but it is a direct check on whether a long context window is actually usable. A perfect score means no query in that run was missed.

Why it matters

The abab models were proprietary: they were offered through the company rather than released as weights others could download and run. That made the abab line a product line as much as a research line, with each version defined by what the company chose to ship.

The series also marks the pace of the line. From abab1 in April 2022 to abab 6.5 in April 2024, MiniMax went through four named generations of text models^[1], and the shift to mixture of experts arrived with abab 6 at the start of 2024^[3].

The distribution model changed later. In January 2025, MiniMax followed the abab line with the open-weight MiniMax-01 series^[1] — a different arrangement, in which the weights themselves are published. Read against that later step, abab 6.5 sits at the end of the period in which MiniMax's language models were closed releases, and it is the version that carried the company's long-context and mixture-of-experts claims into that period's last year.