MiniMax launches new M3 language model at significantly lower cost
Chinese AI startup MiniMax has unveiled its new large language model, M3. The model offers advanced coding and agent capabilities with a million-token context window at a fraction of the cost of competitors.
Chinese AI startup MiniMax released its new large language model, M3, on Sunday. The company claims the model provides advanced coding and agent performance with a one-million-token context window. MiniMax states that M3 outperforms leading proprietary models from Google and OpenAI on certain benchmarks while operating at a significantly lower cost.
The company also announced plans to release the model under an open-source license with "open weights," allowing enterprises to download and customize it for free. This release is expected within the next 10 days. Currently, M3 is available via the MiniMax API at a special discounted introductory price.
At the core of the model's efficiency is its new MiniMax Sparse Attention (MSA) technique, a departure from traditional Transformer architectures. Standard attention mechanisms scale quadratically, making long input processing computationally expensive. MSA reportedly speeds up operations significantly and reduces computational costs.
MiniMax M3's pricing starts below $0.6 per million input tokens and $2.4 per million output tokens at full price. This positioning offers enterprises a substantially more affordable alternative compared to current leading closed-source models, which are priced multiple times higher.