MiniMax Rolls Out Self-Developed "Abab6," China's First Trillion-Parameter MoE Large Language Model, at Full Scale | Oasis Vitality

Recently, MiniMax fully released its large language model Abab6, the first MoE (Mixture-of-Experts) trillion-parameter LLM developed in China. Under the MoE architecture, Abab6 gains the ability to handle complex tasks that come with massive parameters, while also enabling the model to train on sufficient data within a given timeframe, significantly boosting computational efficiency. It addresses the shortcomings of Abab5.5 in scenarios requiring more complex processing and finer-grained output control.

Below is a detailed evaluation and application showcase of Abab6 under the MoE architecture.

Abab6 Evaluation Results

We conducted automated evaluations of abab6, abab5.5, GPT-3.5, GPT-4, Claude 2.1, and the commercial version of Mistral-Medium. Since abab5.5 already performed well on simple tasks, we selected three evaluation methods covering more complex problems:

IFEval: This evaluation primarily tests the model's ability to follow user instructions. We ask the model questions with specific constraints, such as "Using XX as the title, list three specific methods, with each description no longer than two sentences," then count how many responses strictly satisfy the constraints.

MT-Bench: This evaluation measures the model's comprehensive English capabilities. We ask the model questions across multiple categories, including roleplay, writing, information extraction, reasoning, math, coding, and knowledge Q&A. Another large model (GPT-4) scores the responses, and we calculate the average.

AlignBench: This evaluation reflects the model's comprehensive Chinese capabilities, using a similar format to MT-Bench.

As shown, abab6 significantly outperforms our previous-generation model abab5.5 across all three test sets. In instruction following, comprehensive Chinese ability, and comprehensive English ability, abab6 substantially surpasses GPT-3.5. Compared to Claude 2.1, abab6 also has a slight edge in instruction following, comprehensive Chinese ability, and comprehensive English ability. Relative to Mistral's commercial version Mistral-Medium, abab6 outperforms it in instruction following and comprehensive Chinese ability, while matching it in comprehensive English ability.

Abab6 is undergoing continuous training. Those interested in accessing the Abab6 model are welcome to visit the MiniMax Open Platform at api.minimax.chat to register.

Abab6 Application Examples

1. Fun Tutoring

Prompt: Teach the following math problem to a child around 7 years old in a simple, easy-to-understand, and fun way: There are 5 apples in a basket. Xiaoming eats 2 apples, then his mom adds 3 more apples. How many apples are in the basket now? To keep the child engaged, please use SpongeBob's speaking style.

Abab: Hey there, little buddies! It's your best friend, SpongeBob! Today we're gonna play a math game and see what happened to the apples in our basket.

At first, we had 5 apples in our basket, just like treasure in the underwater world. Then our good friend Xiaoming came along and ate 2 apples, like gobbling up 2 pearls from the sea. So now we had 5 minus 2, which equals 3 apples left in the basket.

But don't you worry, because Xiaoming's mom is super kind, and she put 3 more apples in, like finding 3 more pearls! So now our basket has 3 plus 3, which equals 6 apples.

So now there are 6 apples in the basket, like 6 shiny new pearls in our underwater world. Did you all get that? Remember, math is like exploring the ocean floor — always full of surprises!

2. Game Design

Prompt: Create a fictional board game for me. The theme should revolve around Shanghai, and the game must include Shanghai's major tourist attractions, historical culture, and distinctive cuisine. Please format your response as a clear game rulebook, with descriptions of each game component organized in ordered lists.

Abab: Game Name: Magic City Exploration

Game Type: City Building and Cultural Exploration

Game Objective:

Players take on the role of Shanghai urban planners, developing their own city by investing in tourist attractions, historical landmarks, and food districts to attract visitors, ultimately becoming Shanghai's most successful urban planner.

Number of Players: 2-4

Play Time: Approximately 60 minutes

Game Components:

1. Game Board: A map of Shanghai showing major tourist attractions, historical cultural areas, and food districts. The map features different colored zones representing different types of plots.
2. Landmark Cards: Describe famous Shanghai tourist attractions such as the Bund, Yu Garden, Oriental Pearl Tower, etc. Each card shows investment cost and tourist appeal score.
3. Historical Culture Cards: Introduce Shanghai's historical culture, such as the Shanghai Museum, the Site of the First CPC National Congress, etc. Each card shows investment cost and cultural heritage score.
4. Food District Cards: List Shanghai's specialty foods such as xiaolongbao, shengjian, Nanxiang xiaolong, etc. Each card shows investment cost and food score.
5. Development Die: A six-sided die used to determine player turn order and development phases.
6. Currency: Represents each player's funds, used to purchase plots and make investments.
7. Scoreboard: Tracks players' tourist appeal scores, cultural heritage scores, and food scores.
8. Player Tokens: Represent each player's position on the game board.

Game Rules:

At the start of the game, each player selects a colored token and places it at the starting point on the game board.

Players take turns rolling the die, moving their token according to the number rolled, and drawing the corresponding card type based on their location.

Players may choose to purchase drawn landmark cards, historical culture cards, or food district cards, and place them in their urban planning area.

Each card has a corresponding investment cost; players must pay the required currency.

On their turn, players may choose to invest in existing buildings to increase their appeal or cultural heritage scores, or purchase new plots and cards.

The game ends after a predetermined number of rounds.

At game end, players calculate their total score based on the sum of tourist appeal score, cultural heritage score, and food score.

The player with the highest total score becomes Shanghai's most successful urban planner and wins the game.

Strategic Tips:

Balance your investments — don't focus solely on one type of building.

Observe opponents' layouts and try to disrupt them by competing for popular tourist attractions or food districts.

Plan your cash flow wisely to ensure sufficient funds for investments and unexpected events.

Use historical culture cards to enhance your city's cultural depth and attract visitors interested in history and culture.

Magic City Exploration will lead players deep into Shanghai's culture and charm, allowing them to enjoy the game while growing to love and understand Shanghai through engaging city building and economic competition.

3. Instruction Following

Prompt: What story does The Legend of Qin tell?

Answer in two paragraphs; the first paragraph's first sentence must begin with "Hailuo AI," the second sentence with "极," and the third sentence with "智"; the second paragraph must begin with "STARFIELD," and include the keywords "明日" and "梦" at appropriate positions.

Abab: Hailuo AI shells hold secrets of history, and The Legend of Qin uses such a mysterious object as its opening to unfold an adventure across time and space. 极尽 imagination and creativity, this work combines the historical backdrop of late Qin dynasty with wuxia and fantasy elements, telling the story of a group of heroic young men and women who fight for their ideals and beliefs in turbulent times. 智勇双全 protagonists must not only face powerful enemies but also unravel historical mysteries one by one, their adventures filled with challenges and growth.

STARFIELD之下, the dreams of 明日 in The Legend of Qin appear especially precious. The protagonists, caught in the currents of history, continuously pursue their own tomorrow, their stories inspiring viewers to contemplate the relationship between individuals and history, and how to persist in their 梦 through adversity. The dawn of 明日 always seems to emerge at the darkest moments, and The Legend of Qin conveys beautiful aspirations for the future and persistent pursuit of dreams through such storytelling.

4. Meeting Summary

Prompt: We asked abab6 to summarize all topics from a 10-minute meeting with 7 participants based on the [Meeting Content], explicitly noting any [Action] items within the content. The number of topics should not exceed 5. Please output in the following format:

Topic 1: xxxxx

Topic Content: xxxxx

Time: xxxxx

Topic 2: xxxx

Topic Content: xxxx

Time: xxxx

Abab:

Topic 1: Accelerating Project Progress While Ensuring Quality

Topic Content: Speaker 5 emphasized ensuring work quality while accelerating progress; Zhang Hua stated that stability is the prerequisite.

Time: 10:02:15

Topic 2: In-depth Discussion of Data Model

Topic Content: Speaker 8 proposed an in-depth discussion of the data model from yesterday's meeting; Speaker 2 and Speaker 5 indicated they would participate and provide materials and ideas.

Time: 10:02:33

Topic 3: Backup Issue

Topic Content: Speaker 3 asked whether the backup issue mentioned in the last meeting had been resolved; Speaker 6 indicated it was still being handled and expected to be resolved tomorrow.

Time: 10:03:00

Topic 4: Market Research and Questionnaire Development

Topic Content: Speaker 7 asked about the status of the market research questionnaire discussed two days ago; Speaker 2 stated that a draft had been prepared and planned to complete it by the weekend.

Time: 10:03:56

Topic 5: Work Assignment and Skills Summary

Topic Content: Speaker 5 proposed discussing upcoming work assignments to let everyone perform in their strongest areas; Zhang Hua stated he would compile everyone's skills and preferences before finalizing through further discussion.

Time: 10:04:42

5. Write a Poem

Prompt: Write an English acrostic poem, 5 lines total, with the first letters spelling "ABAB6," on the theme of MiniMax's ABAB6 large language model release.

Abab:

Astonishing breakthroughs in AI,

Bringing forth a new era of digital intelligence,

Awesome power in the palm of your hand,

Beyond imagination, the ABAB6 model shines,

Sixth sense in technology, MiniMax leads the way.