AI/TLDR

abab5.5 / abab5.5s

MiniMax's 2023 production chat models that reached GPT-3.5-level quality and powered its early assistant and companion apps.

Overview

abab5.5 and abab5.5s are MiniMax's 2023 production-grade chat models, part of the abab family the Shanghai lab opened to developers through its MiniMax Open Platform (launched April 2023). abab5.5 was the workhorse Chinese-language text model behind MiniMax's early assistant and companion experiences, and contemporaneous user reports placed its writing, chatting, and question-answering quality at roughly the level of OpenAI's GPT-3.5.

abab5.5-chat exposed a combined input-plus-output limit of 16,384 tokens, while abab5.5s-chat was a smaller, lighter variant with an 8,192-token limit aimed at faster, lower-cost interactions. Both were proprietary, API-served models rather than open-weight releases. In the SuperCLUE October 2023 Chinese-LLM leaderboard, MiniMax-Abab5.5 placed around 10th with a total score of 59.57 (48.13 on the open multi-turn track, 76.72 on the objective-question track).

abab5.5 is the model line that powered MiniMax's breakout consumer apps of the period — the Glow chat app and its successors Talkie (international, June 2023) and Xingye / 星野 (China, September 2023). As the earliest abab generation tracked here, it marks the start of MiniMax's publicly documented lineage that later scaled to the 100B+ MoE abab6 (which MiniMax said fixed issues abab5.5 hit in complex scenarios) and the trillion-parameter abab6.5 series. The abab5.5 models have since been retired and no longer appear in MiniMax's current model list; this page preserves them as the foundation of the legacy abab line.

Released2023
LicenseProprietary
WeightsAPI only
Context16K (abab5.5) / 8K (abab5.5s)
Max outputShared with input — 16,384 total tokens (abab5.5), 8,192 total (abab5.5s)
ArchitectureTransformer language model (dense; pre-dates the MoE abab6 generation)
Knowledge cutoff2023 (not formally disclosed)
ModalitiesText
StatusDeprecated — legacy 2023 line, superseded by abab6 (2024) and removed from MiniMax's current model list

Benchmarks

  1. SuperCLUE total (Chinese LLM benchmark, Oct 2023)59.57%
  2. SuperCLUE OPEN (multi-turn open, Oct 2023)48.13%
  3. SuperCLUE OPT (objective questions, Oct 2023)76.72%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

Strengths

  • Among MiniMax's first publicly available production chat models (2023)
  • Reached roughly GPT-3.5-level quality on writing, chat, and Q&A per contemporaneous reports
  • Two tiers: abab5.5 (16K) for quality, lighter abab5.5s (8K) for speed and cost
  • Powered MiniMax's breakout companion apps Talkie and Xingye
  • Top-10 finish among Chinese LLMs on SuperCLUE's October 2023 leaderboard

Best for

  • Conversational assistants and companion-style chat apps (Talkie, Xingye)
  • Lightweight, cost-sensitive chat via the smaller abab5.5s
  • General Chinese-language text generation and dialogue
  • Early API-based application prototyping on the MiniMax Open Platform

How to access

ProviderModel ID
MiniMax ↗abab5.5-chat
MiniMax ↗abab5.5s-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.

VersionReleasedContextLicense
abab6.5 / abab6.5scurrent2024-04-17200KProprietary
abab62024-0132KProprietary
abab5.5 / abab5.5s202316KProprietary

FAQ

What are abab5.5 and abab5.5s?

They are MiniMax's 2023 production chat models in the abab family. abab5.5 is the standard model and abab5.5s is a smaller, lighter variant tuned for faster, lower-cost responses. Both are Chinese-language-focused text chat models served via MiniMax's API.

What were their token limits?

abab5.5-chat supported a combined input-plus-output limit of 16,384 tokens, while the lighter abab5.5s-chat supported 8,192 tokens. These caps cover input and output together rather than a separate larger context window.

How good was abab5.5?

Contemporaneous reports put its writing, chat, and question-answering quality at roughly GPT-3.5 level. On the SuperCLUE October 2023 Chinese-LLM leaderboard, MiniMax-Abab5.5 scored 59.57 overall (about 10th place), with 48.13 on the open multi-turn track and 76.72 on objective questions.

Are the abab5.5 models still available?

No. They were proprietary, API-only models and have since been retired. MiniMax moved on to the MoE abab6 (early 2024) and trillion-parameter abab6.5 (April 2024) lines, and later the MiniMax-01 and open-weight M-series; abab5.5 no longer appears in MiniMax's current model list.