MiniMax M3: Revolutionizing Low-Cost Fast Decoding

On June 1, a Shanghai-based lab quietly introduced a groundbreaking model, the MiniMax M3, capable of decoding a context of 1 million tokens at an impressive speed, 15.6 times faster than its predecessor. What makes this advancement even more remarkable is its cost, which is about 8% of that of Claude Opus for similar performance.

The Innovation Behind MiniMax M3

The key element of this technological advancement lies in the model's architecture, known as MiniMax Sparse Attention (MSA). Unlike standard attention methods, which become costly at such extensive context lengths, MSA employs an optimized approach. It relies on a lightweight indexing branch above the attention by grouped queries to select relevant KV cache blocks, executing attention only on these blocks. This allows for the use of actual, uncompressed key-values while optimizing GPU memory access through a pattern of KV outer gather Q.

Comparison with Other Approaches

The article highlights the differences between MSA and other approaches such as DeepSeek's latent attention (MLA) and native sparse attention (NSA). These comparisons underscore how MSA manages to reduce costs while maintaining high efficiency for large context sizes.

Benchmarks and Reservations

Although the initial benchmarks reported by the provider are promising, the article notes that independent testing could not be conducted at the time of launch, as the model weights were not yet available. While MiniMax M3 shows competitiveness in coding, it is less effective in the multimodal domain and in managing hallucinations.

An Economically Viable Model

Price is a determining factor for MiniMax M3, with extremely low costs per million tokens for input and output, making long-context workflows economically viable. The article's author offers tips for a quick start with M3 via OpenRouter or the MiniMax API, and recommends practical testing to evaluate long-context behavior.

In conclusion, while MiniMax M3 may not be the smartest model overall, its cost and economic viability for processing 1 million tokens open a new category of products. However, uncertainty remains due to the lack of independent benchmarks and the unavailability of model weights.