MiniMax Launches MiniMax-M1: A Long-Context Reasoning LLM from China
Shanghai-based AI startup MiniMax,has officially released MiniMax M1, an open source large language model designed for long context reasoning and agentic capabilities.
With this release, MiniMax aims to challenge Western AI leaders such as OpenAI, Google, and Anthropic by offering performance parity at a fraction of the cost, both in compute and API access.
Key Features
Massive Context Window: MiniMax-M1 supports up to 1 million input tokens and 80,000 output tokens.
This enables the model to process entire codebases, legal documents, or multi-document conversations with high coherence.
Dual Variants: M1 is available in two configurations — M1-40k and M1-80k — with “thinking budgets” of 40K and 80K output tokens.
In most benchmarks, M1-80k outperforms its smaller sibling, particularly in complex reasoning tasks.
Foundation Model & Architecture: Built on MiniMax-Text-01, a 456-billion-parameter foundation model.
Uses a hybrid mixture-of-experts (MoE) setup combined with a proprietary Lightning Attention mechanism — significantly improving memory efficiency, speed, and context handling.
Compute Efficiency: M1 consumes less than half the computing power of DeepSeek-R1 for tasks involving up to 64,000 token generations.
It was trained on 512 H800 GPUs over just 3 weeks, with a total cost of ~$534,700 — a fraction of what most foundation models require.
Benchmark Results
SWE-bench (Software Engineering): M1-40k scored 55.6% and M1-80k scored 56.0% — just behind DeepSeek-R1-0528’s 57.6%, but well ahead of other open-weight models.
Third-party benchmarks show M1 performing on par with leading closed-source models from OpenAI, Google, and Anthropic in: Math (86% AIME 2024), Coding, Domain-specific reasoning, Agentic tasks (TAU-bench, LiveCodeBench, MRCR, SWE-bench).
Agentic Capabilities & Developer Ecosystem
M1 is fully open-source under Apache 2.0 and integrates easily with:
vLLM backend, Hugging Face Transformers, Structured function calling & long-context chat history, The model powers an advanced chatbot API that supports web search, image/video generation, voice cloning, and TTS synthesis.
Pricing
MiniMax is coupling its open-source release with aggressively low-cost API access:
- $0.40 per million tokens for inputs up to 200,000 tokens
- $1.30 per million tokens for the full 1-million-token capability
- Free, unlimited access on the MiniMax app and web platform for personal use.
According to the company, “Due to its relatively efficient use of training and inference computing power, they are offering unlimited free use on the MiniMax APP and Web, and providing APIs on our official website at the industry’s lowest prices.”
Conclusion
MiniMax-M1 combines massive reasoning power, long context capabilities, and unmatched efficiency — all while being truly open-source and budget-friendly. It represents a new generation of high-performing, accessible AI, and a powerful step toward decentralizing global AI innovation.
Baidu Open-Sources Its Ernie AI Model to Compete with OpenAI
Alibaba Launches AI Model Qwen-VLo
Meta, Anthropic Secure Legal Wins in AI Copyright Battles
Microsoft Launches “Mu” – A Smart Local AI for Your PC Settings
Google Launches Personal Styling App “Doppl”
