MiniMax Launches M3 Model With 1M Context Window at $0.30 Per Million Input Tokens

MiniMax has released M3, a multimodal foundation model that processes text, image, and video inputs with a 1-million-token context window. The model costs $0.30 per million input tokens (50% launch discount) and $1.20 per million output tokens (50% launch discount), available through OpenRouter.

Technical Architecture

M3 is built on MiniMax Sparse Attention (MSA), which replaces full attention mechanisms with KV-block selection. According to MiniMax, this reduces per-token compute costs to approximately 1/20th of the previous generation at 1M tokens while maintaining quality across most tasks. The company claims substantially faster prefill and decode speeds compared to traditional full attention models.

The model was trained as a native multimodal system on interleaved data and fine-tuned using an interactive user-simulator framework designed for multi-turn, production-like collaboration.

Target Use Cases

MiniMax positions M3 for:

• Long-horizon agentic workflows requiring sustained context
• Coding tasks with large codebases
• Tool use and function calling
• Multi-step tasks requiring extended reasoning chains

The model outputs text only, despite accepting multimodal inputs.

Pricing Context

At current 50% discounted rates, M3's input pricing of $0.30 per million tokens undercuts most long-context competitors. For comparison:

• Claude 3.5 Sonnet (200K context): $3.00 input / $15.00 output per million tokens
• GPT-4 Turbo (128K context): $10.00 input / $30.00 output per million tokens
• Gemini 1.5 Pro (2M context): $1.25 input / $5.00 output per million tokens (for prompts under 128K)

The model's full-price rates ($0.60 input / $2.40 output) would still position it as cost-competitive for extended context applications.

What This Means

M3 represents MiniMax's entry into the ultra-long-context market with a focus on cost efficiency through architectural innovation. The sparse attention approach directly addresses the computational bottleneck that has made million-token contexts prohibitively expensive for most applications. However, the model's quality at scale and real-world performance on complex multimodal tasks remains to be validated by independent benchmarks. The emphasis on multi-step agentic workflows suggests MiniMax is targeting enterprise automation and developer tooling markets rather than consumer chatbot applications.