MiniMax has quickly emerged as a major player in the competitive AI landscape. It is a Shanghai-based AI foundation model company and is listed on the Hong Kong Stock Exchange in early 2026. MiniMax stands alongside DeepSeek, Anthropic, and OpenAI as a pioneer of frontier-grade artificial intelligence.
What makes MiniMax distinct is its full-stack multimodal focus. While many AI companies specialize in either text or video, MiniMax designs native foundation models across text, code, image, video, and speech, making frontier intelligence available at a fraction of Western proprietary pricing.
This comprehensive guide covers everything you need to know about MiniMax: its flagship models, technical specifications, real-world pricing, competitor comparisons, and a step-by-step example of how to build with it.
What is MiniMax and what models are there?
Founded in 2022 by Yan Junjie (former deputy director of research at Huawei), MiniMax began with the vision to build native multimodal foundation models. Rather than wrapping third-party services, MiniMax trains its models from scratch.
MiniMax divides its core technologies across three main product lines:
1. The M-series (Text, code, & reasoning)
- MiniMax M3: The company’s flagship 428B-parameter open-weight model. According to the official MiniMax M3 release, M3 combines frontier coding performance, a 1-million-token context window, and native multimodality (text, image, and video processing) into a single base model.
- MiniMax M2.7 / M2.7-Highspeed: Prior-generation flagship text models optimized for ultra-fast conversational responses and structured output generation.
2. The Hailuo AI series (Video & visual generation)
- Hailuo 2.3 & Hailuo 2.3 Fast: Generative video systems accessible via the Hailuo AI platform. They produce cinematic 768p and 1080p video clips from text prompts or static images with natural character motion and camera dynamics.
- Hailuo 02: Legacy video model suited for standard definition (512p) and rapid prototyping.
3. Speech-01 & Audio series (Voice synthesis)
- MiniMax Speech-01 / Audio: High-fidelity voice cloning and real-time text-to-speech models capable of generating expressive speech with low latency.
Technical specifications of MiniMax models
MiniMax models are engineered specifically for scale, efficiency, and real-time execution.
Key specifications table
| Feature / Metric | MiniMax M3 (Text/Code) | Hailuo 2.3 Fast (Video) | MiniMax Speech-01 |
|---|---|---|---|
| Primary modality | Text, Code, Image & Video Input | Image-to-Video & Text-to-Video | Text-to-Speech & Voice Cloning |
| Context window / duration | 1,048,576 tokens (~750,000 words) | 6s to 10s video clips (24 fps) | Real-time streaming audio |
| Architecture | ~428B MoE (~23B active) + MSA | Diffusion-Transformer (DiT) | Multi-scale Autoregressive |
| Resolution / quality | N/A | Up to 1080p Native HD | 44.1kHz High-Fidelity Audio |
| SWE-bench pro score | 59.0% | N/A | N/A |
| BrowseComp Score | 83.5 | N/A | N/A |
| Terminal-bench 2.1 | 66.0% | N/A | N/A |
MiniMax Sparse Attention (MSA) architecture
Standard full attention scales quadratically, making a 1M-token context window slow and costly. To solve this, MiniMax published the MiniMax Sparse Attention arXiv paper.
As detailed in the MiniMax-M3 GitHub repository, MSA uses a lightweight Index Branch to scan the entire context and filter Key-Value (KV) blocks. The Main Branch then runs compute solely on those selected blocks. This design achieves:
- 9.7x faster prefill speeds for massive codebase prompts.
- 15.6x faster decoding (~105 tokens/sec) compared to standard full-attention models.
- 28.4x reduction in per-token compute at max context length.
What can you do with MiniMax?
Because of its native multimodal handling and high efficiency, MiniMax enables several demanding AI applications:
- Full-repo code refactoring & auditing: Ingest an entire software repository into a single prompt. M3 analyzes multi-file dependencies, identifies bug origins, and generates fix pull requests without losing context.
- Autonomous agent workflows: Using tool-calling capabilities (scoring 74.2% on MCP Atlas), M3 can execute CLI commands, run web searches, and complete multi-step research autonomously.
- Generative video production: Creators and marketing teams use Hailuo AI to create commercial assets, product visualizers, and social media videos directly from text or images.
- Interactive voice assistants: Integrate low-latency voice synthesis to create human-like conversational agents for customer support or dynamic storytelling.
MiniMax pricing: Subscriptions vs. API costs
MiniMax offers two distinct billing tracks: developer API pricing (charged per token or per generation) and consumer/creator subscriptions.
1. Developer API costs
| Model | Input cost (per 1M) | Output cost (per 1M) | Cached input read |
|---|---|---|---|
| MiniMax M3 (≤512k tokens) | $0.30 | $1.20 | $0.06 |
| MiniMax M3 (>512k tokens) | $0.60 | $2.40 | $0.12 |
| MiniMax M2.7 | $0.30 | $1.20 | $0.06 |
| Hailuo 2.3 Fast (768p / 6s) | $0.19 per video | N/A | N/A |
| Hailuo 2.3 Fast (1080p / 6s) | $0.33 per video | N/A | N/A |
Pro-tip on prompt caching: With a ~90% prompt cache hit rate in multi-turn coding sessions, effective input pricing drops to ~$0.08 per 1M tokens.
2. Monthly subscription plans
- Developer token subscriptions (MiniMax M3):
- Plus ($20/mo): ~1.7 Billion tokens
- Max ($50/mo): ~5.1 Billion tokens
- Ultra ($120/mo): ~9.8 Billion tokens
- Hailuo AI creator subscriptions (Video generation):
- Free ($0/mo): Watermarked outputs, basic queue.
- Standard (~$9.99–$14.99/mo): ~1,000 credits/mo (~20–30 videos), priority queue.
- Pro (~$34.99–$54.99/mo): ~4,500 credits/mo, includes commercial usage rights.
- Max / Ultra (~$124.99–$199.99/mo): Up to 20,000 credits, unlimited generations.
How does miniMax compare to competitors?
Text & coding comparison: MiniMax M3 vs. western flagships
| Metric / Dimension | MiniMax M3 | OpenAI GPT-5.5 | Anthropic Claude Opus 4.8 | Google Gemini 3.1 Pro |
|---|---|---|---|---|
| Context window | 1,048,576 tokens | 1M tokens | 200k – 1M tokens | 1M tokens |
| SWE-bench pro | 59.0% | ~58.6% | ~69.2% | ~54.2% |
| Input price / 1M | $0.30 promo $0.60 std ($0.06 cached) | ~$5.00 ($0.50 cached) | ~$5.00 ($0.50 cached) | $2.00 (≤200k) $4.00 (>200k) |
| Output price / 1M | $1.20 promo ($2.40 std) | ~$30.00 | ~$25.00 | $12.00 (≤200k) $18.00 (>200k) |
| Model weights | Open weights | Proprietary closed | Proprietary closed | Proprietary closed |
Takeaway: MiniMax M3 delivers SWE-bench scores comparable to top-tier closed models while costing 15x to 20x less on input tokens.
Video generation comparison: Hailuo 2.3 vs. Competitors
| Provider | Entry price | Native resolution | Commercial rights | Speed |
| MiniMax Hailuo 2.3 | ~$9.99/mo | 1080p HD | Included in Pro ($34.99+) | Very Fast (Fast Mode) |
| Runway Gen-4 | $12.00/mo | Up to 4K | Standard tier and above | Moderate |
| Kling AI 3.0 | ~$10.00/mo | 1080p | Included | Moderate |
| Luma Dream Machine | $30.00/mo | 1080p | Plus tier ($30+) | Fast |
Example: Building an automated code review agent with MiniMax M3
To illustrate how MiniMax M3 operates in production, here is how to build a Python-based GitHub Code Review Agent using the MiniMax API.
This agent reads a codebase file structure, passes up to 500k tokens of code into MiniMax M3, and requests a comprehensive security and refactoring audit.
Python
import os
import requests
# Set MiniMax API key and endpoint
MINIMAX_API_KEY = os.getenv("MINIMAX_API_KEY")
API_URL = "https://api.minimax.chat/v1/text/chatcompletion_v2"
def audit_codebase(code_content: str):
headers = {
"Authorization": f"Bearer {MINIMAX_API_KEY}",
"Content-Type": "application/json"
}
# Prompting MiniMax M3 using its native tool calling & long-context pipeline
payload = {
"model": "MiniMax-M3",
"messages": [
{
"role": "system",
"content": "You are a Senior Security & Principal Software Architect. Analyze the codebase provided and identify memory leaks, security vulnerabilities, and architectural bottlenecks."
},
{
"role": "user",
"content": f"Here is the entire application codebase:\n\n{code_content}\n\nProvide a structured bug fix report and refactoring plan."
}
],
"temperature": 0.1,
"max_tokens": 4096,
"service_tier": "standard" # Options: standard or priority
}
response = requests.post(API_URL, headers=headers, json=payload)
if response.status_code == 200:
result = response.json()
print("--- MINIMAX M3 AUDIT REPORT ---")
print(result['choices'][0]['message']['content'])
else:
print(f"Error: {response.status_code}, {response.text}")
# Example Usage with a large code context string
sample_codebase = """
// Imagine 200,000 lines of Python / C++ code here...
"""
audit_codebase(sample_codebase)Why MiniMax M3 excels at this:
- Whole-repo visibility: You don’t have to break your code into chunks. The 1M token context processes entire projects simultaneously.
- Ultra-low cost: Auditing a 300,000-token codebase costs approximately $0.09 per run on MiniMax M3, compared to $1.00+ per run on rival models.
Conclusion
MiniMax represents a significant shift in the global AI ecosystem. By introducing MiniMax Sparse Attention (MSA), the company has shown that massive 1M-token context windows can be fast and economically viable for high-volume enterprise production.
Whether you are an engineering team looking to lower coding agent costs or a content creator leveraging Hailuo 2.3 for cinematic AI video, MiniMax offers a compelling combination of top-tier performance and disruptive pricing.
Frequently asked questions (FAQ)
Is MiniMax M3 an open-weight model?
Yes. MiniMax M3 is an open-weight model, allowing developers to run inference locally or host instances on private GPU infrastructure using quants (such as 1-bit or 4-bit GGUFs) via Unsloth or llama.cpp.
Can I use Hailuo AI generated videos for commercial purposes?
Commercial usage rights are included with the Hailuo AI Pro Plan (~$34.99–$54.99/mo) or higher tiers. Videos generated on the Free or Standard plan are intended for personal/non-commercial use.
How does MiniMax keep its API costs so low?
MiniMax uses sparse attention (MSA), which selectively executes compute only on relevant KV cache blocks rather than scanning the entire context window for every token generated. This yields a 28x reduction in per-token compute at 1M token context lengths.
Where can I access the MiniMax API?
You can access MiniMax models directly through the official MiniMax Developer Platform or via major model aggregators like OpenRouter, DeepInfra, and Novita AI.
