Overview
abab6 was MiniMax's flagship large language model released in January 2024, and the company positioned it as the first Mixture-of-Experts (MoE) large language model in China and the first Chinese LLM at over a hundred billion parameters built on an MoE design. Rather than activating every parameter for each token, abab6 routes inputs through a subset of specialized "experts," an approach MiniMax adopted to scale total capacity while controlling inference cost. To ship it, MiniMax developed its own MoE training and inference framework.
MiniMax described abab6 as a clear step up from its previous abab5.5 model, with gains in instruction following and in both Chinese and English comprehension. The company said abab6 surpassed GPT-3.5 on its internal evaluations (and, per contemporaneous reporting, Claude 2.1 on several tests) while still trailing GPT-4. MiniMax evaluated the model on instruction-following and comprehension suites including IFEval, MT-Bench, and AlignBench, though it published the comparison only as a chart rather than as a table of exact scores.
abab6 was a hosted, closed model on the MiniMax open platform and was used by the company's enterprise customers in China. It was short-lived as a flagship: in April 2024 MiniMax released the abab6.5 series (abab6.5 and the faster abab6.5s) with a 200k-token context window, and the company later moved to its M-series models. abab6 is no longer listed among the models in MiniMax's current API, making it a legacy entry in the abab line.
| Released | 2024-01-16 |
|---|---|
| License | Proprietary (MiniMax). Offered only as a hosted API on the MiniMax open platform; no public terms beyond standard API usage. |
| Weights | API only |
| Parameters | Not officially disclosed. MiniMax described abab6 as the first Chinese large language model with a parameter scale of over a hundred billion built on a MoE (Mixture-of-Experts) architecture; an exact count was never published. |
| Context | 32K |
| Architecture | Dense-to-sparse Mixture-of-Experts (MoE) transformer. Model parameters are split across multiple "experts," and only a subset of experts is activated per token at inference, which lets the model carry a large total parameter budget while keeping per-token compute lower. MiniMax built its own MoE training and inference framework for abab6. |
| Knowledge cutoff | Not officially disclosed by MiniMax. |
| Modalities | text |
| Status | Legacy / retired. Released January 2024 and superseded by the abab6.5 series (April 2024) and later MiniMax M-series models; abab6 is no longer listed among the models in MiniMax's current API. |
Strengths
- Pioneered MoE architecture in China — the first domestic large language model to ship on a Mixture-of-Experts design at over 100B total parameters
- Stronger instruction following and Chinese/English comprehension than MiniMax's earlier abab5.5
- Outperformed GPT-3.5 on MiniMax's internal evaluations
- Backed by MiniMax's in-house MoE training and inference framework, built specifically for this model line
Best for
- Chinese-language assistants and chatbots (its core target market)
- Instruction-following tasks such as structured answers and content generation
- Bilingual (Chinese/English) comprehension and writing
- Powering consumer and enterprise products on the MiniMax open platform (now better served by abab6.5 or the M-series)
How to access
| Provider | Model ID |
|---|---|
| MiniMax ↗ | abab6-chat |
abab (legacy) — every version
The full lineage of the abab (legacy) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| abab6.5 / abab6.5scurrent | 2024-04-17 | 200K | Proprietary |
| abab6 | 2024-01 | 32K | Proprietary |
| abab5.5 / abab5.5s | 2023 | 16K | Proprietary |
FAQ
What is MiniMax abab6?
abab6 was MiniMax's flagship large language model, released in January 2024. MiniMax called it the first Mixture-of-Experts (MoE) large language model in China and the first Chinese LLM at over a hundred billion parameters on an MoE architecture.
Is abab6 still available?
No. abab6 is a legacy model. It was superseded by the abab6.5 series in April 2024 and later by MiniMax's M-series, and it is no longer listed among the models in MiniMax's current API.
How did abab6 compare to GPT-4 and GPT-3.5?
MiniMax said abab6 surpassed GPT-3.5 on its internal evaluations (and, per contemporaneous reporting, Claude 2.1 on several tests) but still trailed GPT-4. It was assessed on suites including IFEval, MT-Bench, and AlignBench, though MiniMax published the comparison only as a chart, not as exact scores.
Is abab6 open-weights?
No. abab6 was a proprietary, closed model offered only through MiniMax's hosted API; its weights were never released.