MiniMax: The Chinese AI Lab Behind MiniMax-Text-01 and Hailuo Video

MiniMax is a Chinese AI lab founded in 2021, backed by Alibaba and Tencent. It operates two distinct product lines: MiniMax-Text for language models and Hailuo AI for video generation. In 2026 it has emerged as one of the more capable and cost-efficient providers in the crowded Chinese LLM space.

MiniMax models

MiniMax-Text-01

MiniMax's flagship language model — a 456B parameter mixture-of-experts model with a 1M token context window. It's one of the longest-context models publicly available via API.

MiniMax-VL-01

The vision-language variant — accepts images alongside text. Same 1M context window as Text-01, with strong performance on document understanding and visual reasoning tasks.

MiniMax vs the Chinese LLM field

MiniMax's main differentiator is the 1M context window at $0.20/1M input — competitive with Gemini 2.5 Flash on context length and cheaper than Gemini on output pricing. DeepSeek V4 Flash is still cheaper overall and has open weights, but MiniMax-Text-01 has stronger scores on long-context retrieval benchmarks.

Hailuo AI — video generation

MiniMax's video product, Hailuo AI, has become one of the most-used AI video generators globally — competing with Sora, Kling, and Runway. It supports text-to-video and image-to-video generation at resolutions up to 1080p. This is a separate product from the LLM API and is accessed via hailuoai.com.

The Lightning Attention architecture

MiniMax-Text-01 uses a hybrid attention architecture combining standard softmax attention for short sequences with "Lightning Attention" — a linear attention variant — for long contexts. This is how the 1M context window remains computationally feasible: linear attention scales O(n) rather than O(n²) with sequence length.

The same pattern appears in other long-context models (Gemini uses a different approach, SSMs). It's worth understanding because it affects what kind of long-context tasks the model handles well — retrieval and summarization perform better than tasks requiring dense cross-document reasoning.

Data privacy

Like DeepSeek and Kimi, MiniMax API calls go to China-based servers. For sensitive workloads, the open-weight models (DeepSeek, Qwen) that can be self-hosted are a better fit.

Using the MiniMax API

MiniMax's API is OpenAI-compatible at https://api.minimax.chat/v1:

from tokoscope import wrap
from openai import OpenAI

client = wrap(OpenAI(
    base_url="https://api.minimax.chat/v1",
    api_key="YOUR_MINIMAX_KEY"
), api_key="ts_live_...")

response = client.chat.completions.create(
    model="MiniMax-Text-01",
    messages=[{"role": "user", "content": "Hello"}]
)